Internet data center fault detection method based on artificial intelligence

Abstract

The invention relates to the field of electric digital data processing, in particular to an Internet data center fault detection method based on artificial intelligence, which comprises the following steps of: acquiring monitoring data of an Internet data center, marking the state of the monitoring data, calculating the dynamic sampling probability of each sample, constructing the space-time correlation tensor of each sample, and calculating the fault of the Internet data center according to the space-time correlation tensor. A weighted time sequence data matrix of each sample is generated, a fault classification probability is generated, a dynamic category weight coefficient of each category is calculated, and an adaptive loss function is calculated; calculating a layering gradient; and performing automatic fault diagnosis on newly collected real-time monitoring data by using the trained fault detection model. An existing detection method has the problem of poor adaptability, and the method provided by the invention is good in adaptability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electric digital data processing, in particular to an internet data center fault detection method based on artificial intelligence. BACKGROUND

[0002] More and more enterprises and units begin to handle business through the Internet. Since the construction platform needs to build a machine room, build a system, and hire development and maintenance personnel, not only a large amount of capital investment is needed, but also it is difficult to achieve professional service quality. Therefore, enterprises and units mostly choose to use the Internet basic platform services and various value-added services provided by the Internet data center. The Internet data center carries the core tasks of user data storage, business system operation, etc. in key fields such as finance, e-commerce, government affairs, and medical treatment. If the Internet data center fails and is not handled in time, it will cause direct business interruption, data loss risk, and other chain reactions. In order to avoid this situation, people take the mode of passive repair and active prevention to cope with it. Active prevention means long-term monitoring of the Internet data center by using the Internet data center fault detection method, and analyzing the monitoring data to discover faults in advance.

[0003] The existing detection method mostly relies on simple rule-based judgment and manual intervention, and it is difficult to cope with real-time monitoring of large-scale data and identification of complex fault modes, so the existing detection method has the problem of poor adaptability. SUMMARY

[0004] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art, and to provide an Internet data center fault detection method based on artificial intelligence with good adaptability.

[0005] To solve the above technical problems, the present application provides an Internet data center fault detection method based on artificial intelligence, which comprises the following steps:

[0006] S1. Collecting monitoring data of the Internet data center and marking the state of the monitoring data;

[0007] S2. Calculating the dynamic sampling probability of each sample based on the fault sensitivity score of each sample;

[0008] S3. Constructing a spatiotemporal correlation tensor of each sample, and decomposing the constructed spatiotemporal correlation tensor of each sample into a product form of a core tensor and a plurality of factor matrices;

[0009] S4. Generating a weighted time series data matrix of each sample based on the normalized importance weight of each dimension index of each sample;

[0010] S5. Extracting sequence dynamic features, extracting index correlation features, performing double-channel feature fusion, and generating a fault classification probability;

[0011] S6. Calculate the dynamic category weight coefficient of each category, and calculate the adaptive loss function;

[0012] S7. Calculate the hierarchical gradient, perform the fault sensitivity oriented gradient clipping, and perform the hierarchical parameter update and non-negative projection;

[0013] S8. Use the trained fault detection model to automatically diagnose the newly collected real-time monitoring data.

[0014] As a further improvement of the present application: the steps of S1 include:

[0015] S101. Deploy multiple types of sensors at key positions of the Internet data center to continuously collect monitoring data of key performance indicators and environmental parameters that can reflect the system running state;

[0016] S102. According to the running state, the time period of each sample is marked.

[0017] As a further improvement of the present application: the steps of S2 include:

[0018] S201. Statistically, all index dimensions of each sample exceed the three standard deviation threshold of the index at all time points. The absolute deviation value is obtained by accumulating all the absolute deviation values, and the fault sensitivity score of each sample is obtained;

[0019] S202. Based on the calculated fault sensitivity scores of all samples, the Softmax function is used for normalization processing, and the sampling probability of each sample is calculated.

[0020] As a further improvement of the present application: the steps of S3 include:

[0021] S301. Perform outer product operation on the original time series data matrix of each sample to generate a three-dimensional spatiotemporal correlation tensor of each sample;

[0022] S302. Decompose the constructed spatiotemporal correlation tensor of each sample into the product form of the core tensor and multiple factor matrices.

[0023] As a further improvement of the present application: the steps of S4 include:

[0024] S401. The core tensor of each sample is flattened into a one-dimensional vector and input into a multi-layer perception network, and the multi-layer perception network outputs the original importance score vector of each sample;

[0025] S402. Convert the original importance score vector of each sample into the normalized importance weight of each dimension index of each sample through the Softmax function with temperature coefficient;

[0026] S403. Multiply the normalized importance weight of each dimension indicator of each sample with the original time series data matrix element by element to generate a weighted time series data matrix of each sample.

[0027] As a further improvement of the present application: the step S5 comprises:

[0028] S501. Input the weighted time series data matrix of each sample into the gated recurrent unit network based on the correlation perception, process the entire time series, and take the hidden state vector output at the last time step of the network as the sequence dynamic feature vector;

[0029] S502. Flatten the core tensor obtained by non-negative Tucker decomposition into a one-dimensional vector and input it into a multi-layer perception network to map it to a vector of the same dimension as the sequence dynamic feature vector to obtain an indicator correlation feature vector;

[0030] S503. Perform a splicing operation on the extracted sequence dynamic feature vector and the indicator correlation feature vector to generate a fusion feature vector;

[0031] S504. Input the fusion feature vector into a fully connected layer and process it through a Softmax activation function to output a probability distribution vector of the fault categories of each sample.

[0032] As a further improvement of the present application: the step S6 comprises:

[0033] S601. For each fault category, calculate the dynamic weight coefficient of each category according to the frequency of each category appearing in the training set and the artificially preset cost factor;

[0034] S602. Combine the calculated dynamic weight coefficient of each category and the focal loss mechanism to construct an adaptive loss function.

[0035] As a further improvement of the present application: the step S7 comprises:

[0036] S701. Calculate the gradient of the total loss function value calculated for the current batch of samples with respect to the gated recurrent unit parameter set, the non-negative Tucker decomposition parameter set, and the fully connected classification layer parameter set, respectively;

[0037] S702. Calculate the average sampling probability of the current batch and dynamically adjust the threshold value of gradient clipping using the average sampling probability;

[0038] S703. Use the gradient clipping threshold to clip the three types of parameter gradients of the gated recurrent unit parameters, the non-negative Tucker decomposition parameters, and the fully connected classification layer parameters, respectively, to limit the gradient value within a certain range.

[0039] The beneficial effects of the present application are as follows: first, the method overcomes the problem that random sampling or simple oversampling / undersampling technology cannot effectively solve the class imbalance problem by adopting a fault sensitivity weighted sampling function, dynamically adjusting the sampling probability of the sample based on the fault sensitivity score, and improving the detection ability of the fault sample, especially for the rare fault type.

[0040] Secondly, the method overcomes the problem that the traditional dimension reduction method cannot effectively capture the nonlinear dynamic interaction between the monitoring indicators of the Internet data center by constructing a spatio-temporal correlation tensor, modeling the instantaneous correlation between the monitoring indicators, and extracting the core features with physical interpretability by using non-negative Tucker decomposition, which can more effectively capture the nonlinear dynamic interaction relationship across indicators and avoid the loss of fault mode information.

[0041] Thirdly, the method overcomes the problem that the fault detection model may not respond to key fault features by manually setting the weight of the indicator for fault diagnosis, so that the fault detection model can adaptively focus on the key indicator features when the fault occurs, rather than relying on the fixed weight set by artificial;

[0042] Finally, the method overcomes the problem that the existing technology usually relies on a single feature extraction path and cannot effectively combine recent mutation features and long-term correlation features by extracting sequence dynamic features and indicator correlation features in parallel and combining a dual-channel feature fusion mechanism, which comprehensively considers recent mutations and long-term correlation patterns, significantly improving the detection accuracy and response speed for complex fault patterns.

[0043] In summary, the response of the fault detection model to minority class faults and complex faults is significantly improved, making the method more adaptable to different fault types. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 Performance bar chart of different sampling methods on various faults;

[0045] Figure 2 Dynamic importance sampling probability distribution diagram;

[0046] Figure 3 Multi-index correlation tensor decomposition result diagram;

[0047] Figure 4 Core tensor feature distribution diagram;

[0048] Figure 5 Indicator importance weight dynamic change diagram over time;

[0049] Figure 6Performance comparison chart of different feature fusion methods;

[0050] Figure 7 Two-channel feature space distribution chart;

[0051] Figure 8 Overall performance comparison chart of different fault detection methods;

[0052] Figure 9 False positive rate trend chart over time;

[0053] Figure 10 Performance experiment chart under different load conditions. DETAILED DESCRIPTION

[0054] The application provides an internet data center fault detection method based on artificial intelligence, which comprises the following steps:

[0055] S1. Collecting monitoring data of the internet data center and marking the state of the monitoring data;

[0056] S101. Deploying multiple types of sensors at key positions of the internet data center to continuously collect monitoring data of key performance indicators and environmental parameters that can reflect the system running state;

[0057] The key positions of the internet data center include CPU, memory module, network equipment, disk storage equipment, computer room space and equipment periphery, etc.

[0058] The multiple types of sensors include performance monitoring sensors, network flow sensors, I / O monitoring sensors and temperature sensors.

[0059] The collection is periodic sampling at a fixed time frequency (such as every minute) to form a multi-dimensional time series data stream, and the real-time aggregation of full cluster coverage is realized through a distributed log system;

[0060] The monitoring data includes five index dimensions, specifically: CPU load, memory usage, network flow, disk I / O and temperature; these indicators correspond to key fault perception dimensions such as computing resources, storage performance, communication state and hardware environment, representing key performance and environmental parameters of the internet data center, and are used to quantify abnormal behavior in fault detection;

[0061] The CPU load is obtained by deploying performance monitoring sensors at the CPUs of the Internet data center; the memory usage is obtained by deploying performance monitoring sensors at the memory modules of the Internet data center; the network traffic is obtained by deploying network traffic sensors at the network devices (such as switches and routers) of the Internet data center; the disk I / O is obtained by deploying I / O monitoring sensors at the disk storage devices of the Internet data center; and the temperature is obtained by deploying temperature sensors around the room space and devices of the Internet data center.

[0062] S102. In combination with the operation and maintenance event log, the alarm record and the manual checking report, the time period of each sample is marked according to the running state to form a structured training data set containing time dimension, index dimension and category label;

[0063] The categories of the state marking include disk failure, memory leakage, network interruption, CPU overload, temperature anomaly and normal;

[0064] In the marking process, the sensor data timestamp and the fault event period need to be strictly aligned to ensure the spatio-temporal consistency of the abnormal pattern and the label,

[0065] S2. The dynamic sampling probability of each sample is calculated based on the fault sensitivity score of each sample;

[0066] S201. The absolute deviation value of each sample of all index dimensions at all time points exceeding the three standard deviation threshold of the index is counted, and all the obtained absolute deviation values are accumulated to obtain the fault sensitivity score of each sample, which is used to quantify the fault risk of each sample;

[0067] The calculation formula of the fault sensitivity score F(X i ) of the i-th sample is:

[0068]

[0069] In the formula, F(·) is the fault sensitivity score function, and the larger the output value is, the higher the fault risk is; X i is the i-th sample, which is in the format of a time series data matrix, with dimensions of TxD, T being the number of time steps and D being the number of index dimensions; i is the sample index, used to traverse all samples, with a value range of 1 to N, N being the total number of samples in the training set; d is the index of the index dimension, with a value range of 1 to D, corresponding to different monitoring indexes; t is the index of the time step, with a value range of 1 to T, representing the time point of data collection; I(·) is an indicator function, outputting 1 when the condition is met, and 0 otherwise; The condition of the term is used to determine whether the value exceeds the 3 standard deviation threshold; ​is the indicator value of the ith sample at the dth dimension at the tth time; μ d is the mean of the dth dimension indicator on the training set; σ d is the standard deviation of the dth dimension indicator on the training set; is the absolute deviation value, which calculates the absolute distance between the sample point and the mean.

[0070] S202. Based on the calculated fault sensitivity scores of all samples, the Softmax function is used for normalization processing, and the sampling probability of each sample is calculated. The response intensity of high-score samples in the normalization process is controlled through the fault sensitivity coefficient, so as to adjust the sensitivity of sampling to abnormal values;

[0071] The sampling probability P of the ith sample sample (X i ) is calculated as follows:

[0072]

[0073] In the formula, P sample (·) is the sample sampling probability function, the value range is [0, 1], and the higher the probability indicates the higher priority of the sample being sampled; exp(·) is the natural exponential function, which is used to amplify the influence of high-score samples; λ w is the fault sensitivity coefficient, which controls the response intensity to abnormal values, and λ w may be set to 1.5; j is the sample index different from i, which is used to traverse all samples, and the value range is 1 to N; N is the total number of samples in the training set; X j is the jth sample, which is in the format of a time series data matrix, and the dimension is TxD.

[0074] The fault sensitivity score function is used to quantify the fault risk of the sample, so that the sampling probability is positively correlated with the fault possibility, thereby relieving the class imbalance problem, while avoiding destroying the time series continuity.

[0075] The implementation of the sample sampling probability function is in the format of the Softmax function, and the normalization processing makes the sampling probability positively correlated with the fault sensitivity score.

[0076] The monitoring data of the Internet data center has high dimensions, variable time series length, and sparse fault samples. Conventional random sampling can ignore key fault fragments, leading to underfitting of the fault detection model to the minority class fault mode, and oversampling techniques can easily destroy the time series continuity, while undersampling can lose key information. S2 uses a fault sensitivity weighted sampling function to dynamically improve the weight of samples containing potential fault fragments, in order to relieve the class imbalance problem and preserve the time series integrity.

[0077] S3. Constructing the spatiotemporal correlation tensor of each sample, and decomposing the constructed spatiotemporal correlation tensor of each sample into a product form of a core tensor and multiple factor matrices;

[0078] S301. Performing an outer product operation on the original time series data matrix of each sample to generate a three-dimensional spatiotemporal correlation tensor of each sample, which represents the instantaneous correlation between each monitoring index of the sample at a time point;

[0079] The spatiotemporal correlation tensor of the i-th sample represents the dynamic interaction relationship between the indices, and has a dimension of DxDxT. The spatiotemporal correlation tensor J i of the i-th sample is calculated according to the following formula:

[0080]

[0081] In the formula, X is the transpose of the i-th sample X i , and has a dimension of DxT; is an outer product operator.

[0082] S302. Decomposing the constructed spatiotemporal correlation tensor of each sample into a product form of a core tensor and multiple factor matrices, and imposing a non-negative constraint condition during the decomposition process, requiring the elements of the core tensor and all factor matrices to be non-negative, so as to ensure that the features obtained by the decomposition have physical interpretability. The target of the decomposition is to minimize the difference between the reconstructed tensor and the original tensor;

[0083] [G i ; U (1) , U (2) , U (3) ] is the complete output of the non-negative Tucker decomposition, which is a conventional representation in the field of tensor decomposition, that is, a set of optimal G i , U (1) , U (2) , U (3) is found by minimizing optimization. This set of optimal solutions needs to satisfy two conditions. One is to minimize the Frobenius norm between the tensor product reconstructed by them and the original tensor, and the other is that the elements in the three factor matrices U (1) , U (2) , U (3) must be non-negative;

[0084] The calculation formula of [G i ; U (1) , U (2) , U (3) ] is:

[0085]

[0086] In the formula, G iLet G be the core tensor of the i-th sample, with dimension R1xR2xR3, and define R k Let R be the rank of the k-th mode, R1 be the rank of the 1st mode, R2 be the rank of the 2nd mode, R3 be the rank of the 3rd mode, and R1<<D, R2<<D, R3<<T; U (1) Let U1 be the 1st mode factor matrix with dimension DxR1; U (2) Let U2 be the 2nd mode factor matrix with dimension DxR2; U (3) Let U3 be the 3rd mode factor matrix with dimension TxR3; argmin represents that the optimization objective is to minimize the function, represents that in the optimization process, the values of the core tensor G and the factor matrix U are found to minimize the objective function ‖·‖ F is the Frobenius norm; G is the core tensor variable to be optimized in the decomposition; x k represents the product of the tensor and the matrix along the k-th dimension, that is, x1 represents the product of the tensor and the matrix along the 1st dimension, x2 represents the product of the tensor and the matrix along the 2nd dimension, and x3 represents the product of the tensor and the matrix along the 3rd dimension; U (k) is the factor matrix of the k-th mode, k is the mode dimension index, and k takes values in the range of k = 1, 2, 3; U is the factor matrix, including U (1) , U (2) , and U (3) ;

[0087] G≥0, U (k) ≥0 are non-negative constraint conditions to ensure that the decomposition result has physical interpretability;

[0088] In the optimization process of minimizing , the constraint conditions are G≥0 and U (k) ≥0, so as to realize the non-negative constraint of the core tensor G i , and model the instantaneous correlation between indicators through the spatiotemporal correlation tensor J i , effectively capturing the cross-indicator dynamic interaction features.

[0089] There are nonlinear coupling relationships between the monitoring indicators of the Internet data center, such as the rise of CPU load causing temperature rise. Directly splicing features will ignore the cross-indicator dynamic interaction mode. Conventional linear dimension reduction techniques cannot effectively capture such complex correlation features, resulting in loss of key fault mode information. S3 explicitly models the instantaneous correlation between indicators by constructing a spatiotemporal correlation tensor, and extracts interpretable core features using non-negative Tucker decomposition.

[0090] S4. Based on the normalized importance weight of each dimension indicator of each sample, a weighted time series data matrix of each sample is generated;

[0091] S401. The core tensor of each sample is flattened into a one-dimensional vector and input into a multi-layer perception network, and the multi-layer perception network outputs the original importance score vector of each sample, each element of the original importance score vector of each sample corresponds to a monitoring index, and the element value represents the initial importance of each index;

[0092] The dimension of the original importance score vector of the i-th sample is D, and the original importance score vector of the i-th sample s i The calculation formula is:

[0093] s i = MLP(vec(G i )),

[0094] In the formula, MLP(·) is a multi-layer perception network, and the specific structure is two fully connected layers; vec(·) is a tensor vectorization operation, which converts a multi-dimensional tensor into a one-dimensional vector;

[0095] S402. The original importance score vector of each sample is converted into the normalized importance weight of each dimension index of each sample by a Softmax function with a temperature coefficient;

[0096] The value range of the normalized importance weight of the d-th dimension index of the i-th sample is [0, 1], and the calculation formula of the normalized importance weight w i,d of the d-th dimension index of the i-th sample is:

[0097]

[0098] In the formula, s i,d is the d-th element of the original importance score vector s i ; τ is a temperature coefficient, which controls the sharpness of the weight distribution, for example, τ = 0.5; d ′ is an index distinguishing the index dimension from d, and the value range is 1 to D, corresponding to different monitoring indexes; d is an index of the index dimension, and the value range is 1 to D, corresponding to different monitoring indexes; s i,d′ is the d i -th element of the original importance score vector s ′ ;

[0099] The temperature coefficient τ is used to control the sharpness of the weight distribution, and the smaller the temperature coefficient value, the more concentrated the weight distribution is on a few important indexes; the larger the temperature coefficient value, the more uniform the weight distribution is; when the temperature coefficient τ→0 + , it represents that the weight distribution tends to one-hot form, and when τ→+∞, it represents that the weight distribution tends to uniform distribution.

[0100] S403. Multiply the normalized importance weight of each dimension indicator of each sample with the original time series data matrix element by element to generate a weighted time series data matrix of each sample, so that the fault detection model can adaptively focus on the key indicator features identified in the current sample;

[0101] The dimension of the i-th sample weighted time series data matrix is TxD, and the i-th sample weighted time series data matrix The calculation formula is:

[0102]

[0103] In the formula, X i is the original time series data matrix with a dimension of TxD; and is the Hadamard product; w i is the importance weight vector with a dimension of Dx1;

[0104] The calculation formula of the importance weight vector w i is:

[0105] w i = [w i,1 , w i,2 , …, w i,d , …, w i,D ] T ,

[0106] In the formula, w i,1 is the normalized importance weight of the 1st dimension indicator of the i-th sample; w i,2 is the normalized importance weight of the 2nd dimension indicator of the i-th sample; and w i,D is the normalized importance weight of the Dth dimension indicator of the i-th sample.

[0107] The importance weight vector w i is dynamically generated by the core tensor G i , so that the fault detection model can adaptively focus on the key indicators, such as the sudden change of the disk I / O before the fault, overcoming the limitations of manually setting the weight.

[0108] The contribution of different monitoring indicators of the Internet data center to the fault is significantly different, such as the disk error which is more indicative of the fault than the temperature fluctuation, and the key indicators depend on the specific fault type. The conventional method of manually setting the weight cannot adapt to the dynamic scene change, resulting in insufficient response of the fault detection model to the key fault features. S4 dynamically calculates the importance weight of each indicator based on the correlation pattern between indicators through a learnable importance weight generator, so that the fault detection model can adaptively focus on the key fault features.

[0109] S5. Extract sequence dynamic features and indicator correlation features, perform dual-channel feature fusion, and generate fault classification probability;

[0110] S501. Input the weighted sample time series data matrix into the Gated Recurrent Unit network based on the correlation perception, process the entire time series, and take the hidden state vector output at the last time step of the network as the sequence dynamic feature vector, thereby capturing the recent dynamic changes and mutation features in the input time series data;

[0111] The dimension of the sequence dynamic feature vector is d h The calculation formula of the sequence dynamic feature vector h seq

[0112]

[0113] In the formula, GRU(·) is the Gated Recurrent Unit based on the correlation perception; is the weighted time series data matrix of the i-th sample, representing the instantaneous weighted input of the i-th sample at time point t, and the dimension is TxD;

[0114] S5011. Based on the hidden state vector at the previous time and the weighted input vector at the current time, the original reset gate vector at each time is calculated through a fully connected layer and a Sigmoid activation function, and the original reset gate vector at each time controls the degree of historical state information being retained;

[0115] The dimension of the original reset gate vector at the t-th time is d h The calculation formula of the original reset gate vector h at the t-th time is:

[0116]

[0117] In the formula, Sig(·) is a Sigmoid activation function that compresses the output to the interval [0, 1]; W r is a reset gate weight matrix with a dimension of d h x (d h +D), which is a learnable parameter, d h is the dimension of the hidden state; [·; ·] is a vector concatenation operation, and the output dimension is d h +D; h t-1 is the hidden state vector at the t-1-th time, with a dimension of d h ; is the weighted time series data vector at the t-th time, is a row vector at a specific time step t, with a dimension of 1xD; b r is a reset gate bias vector with a dimension of d h , which is a learnable parameter;

[0118] The calculation formula of the weighted time series data vector h at the t-th time is: ​

[0119]

[0120] In the formula, [t, :] represents a slicing operation, that is, a row vector composed of all index values at the t-th moment;

[0121] S5012. The original reset gate vector at each moment is used to filter the hidden state vector at the previous moment, the filtering result is spliced with the weighted input vector at the current moment, the spliced vector is passed through a fully connected layer and a GELU activation function to generate an intermediate feature vector, the intermediate feature vector is linearly transformed and a correlation attention vector is calculated through a Softmax function, thereby representing the strength of statistical correlation between indicators in the current input;

[0122] The dimension of the intermediate feature vector at the t-th moment is d c , used to carry correlation information, and the calculation formula of the intermediate feature vector c t at the t-th moment is:

[0123]

[0124] In the formula, GE(·) is a GELU activation function, which enhances the nonlinear expression ability; W c is a feature transformation weight matrix, with a dimension of d c ×(d h +D), which is a learnable parameter; b c is an attention bias vector, with a dimension of d c , which is a learnable parameter;

[0125] The dimension of the correlation attention vector is d k , used to represent the strength of statistical correlation between indicators, and the calculation formula of the correlation attention vector k t is:

[0126]

[0127] In the formula, soft(·) is a normalized exponential function, outputting a probability distribution; W k is an attention weight matrix, with a dimension of d k ×d c , which is a learnable parameter, d c is the intermediate feature dimension; d k is the attention vector dimension;

[0128] Taking the square root of the attention vector dimension gives representing a scaling factor to prevent gradient vanishing caused by excessive dot product.

[0129] S5013. Based on the correlation attention vector, the index covariance matrix and the weighted time series data vector, the correlation correction term is obtained. The correlation correction term is multiplied by the correlation enhancement coefficient and then superimposed on the original reset gate vector to obtain the corrected reset gate vector.

[0130] The dimension of the corrected reset gate vector at time t is d. h The corrected and reset gate vector r at time t t The calculation formula is:

[0131]

[0132] In the formula, λ ca To determine the correlation enhancement coefficient and control the correction intensity, λ can be set. ca =0.7; For k t The transpose of the y-axis has a dimension of 1×d. k Σ is the index covariance matrix with dimension D×D, which is pre-computed on the self-training set and then frozen. The term represents the correlation correction term, quantifying the statistical correlation between the current input and historical states, through... The model models the statistical correlation between the current input and historical indicators, corrects the original reset gate, and thus improves the ability to respond to correlational mutations, such as improving the ability to respond to correlational mutations such as network traffic surges accompanied by CPU anomalies.

[0133] The index covariance matrix is ​​calculated using index data from the entire training set. The covariance is calculated for each index dimension. Once the calculation is complete, the index covariance matrix Σ is fixed during the training of the fault detection model and is no longer updated. It is used as a constant matrix to avoid bias caused by changes in statistical characteristics during training.

[0134] S5014. Based on the hidden state vector of the previous time step and the weighted input vector of the current time step, an update gate vector is calculated through a fully connected layer and a sigmoid activation function. The update gate vector controls the degree of update of the old state by the new state. The hidden state vector of the previous time step is filtered by the modified reset gate vector. The filtering result is concatenated with the weighted input vector of the current time step. The candidate hidden state vector is calculated through a fully connected layer and a hyperbolic tangent activation function. The hidden state vector of the current time step is updated by using the update gate vector and combining the hidden state vector of the previous time step and the candidate hidden state vector.

[0135] The dimension of the updated gate vector at time t is d. h The update gate vector z at time t is used to control the degree of state update. t The calculation formula is:

[0136]

[0137] where W z is the updated gate weight matrix with dimension d h ×(d h +D), and b z is the updated gate bias vector with dimension d h , are learnable parameters.

[0138] The dimension of the candidate hidden state vector at the t-th time step is d h , and the candidate hidden state vector at the t-th time step h is calculated as follows:

[0139]

[0140] where tanh(·) is the hyperbolic tangent activation function with an output range of [-1, 1]; W h is the candidate state weight matrix with dimension d h ×(d h +D), and b h is the candidate state bias vector with dimension d h .

[0141] The dimension of the hidden state vector at the t-th time step is d h , and the calculation formula of the hidden state vector h t at the t-th time step is as follows:

[0142]

[0143] The hidden state vector at the last time step is taken as the representation of the entire sequence to capture the global temporal dynamic characteristics, i.e., h T =h seq , and h T is the hidden state vector at the final time step t=T with dimension d h .

[0144] Since the Internet data center fault detection needs to capture long-range temporal dependencies, such as gradual trends hours before the fault, while suppressing instantaneous noise fluctuations, the update gate of the standard Gated Recurrent Unit is sensitive to sudden noise and tends to forget important states, and cannot explicitly model the statistical association between indicators, resulting in insufficient response to associated mutations, such as network traffic surges accompanied by CPU abnormalities; the association-aware Gated Recurrent Unit adopted in S501 uses an association-aware gating mechanism in the Gated Recurrent Unit, which corrects the gating signal by modeling the statistical association between the current input and historical indicators, thereby enhancing the response capability to associated mutations.

[0145] S502. The core tensor obtained by the non-negative Tucker decomposition is flattened into a one-dimensional vector and input into a multi-layer perception network to be mapped to a vector with the same dimension as the sequence dynamic feature vector, obtaining an index correlation feature vector, which has features extracted from the spatiotemporal correlation tensor and reflects long-term correlation patterns between indexes;

[0146] The dimension of the index correlation feature vector is d h , which is used to carry long-term correlation patterns, and the calculation formula of the index correlation feature vector h core is:

[0147] h core = MLP (Flatten (G i )),

[0148] In the formula, Flatten(·) is a tensor flattening operation that converts a three-dimensional tensor into a one-dimensional vector;

[0149] S503. The extracted sequence dynamic feature vector and the index correlation feature vector are spliced to generate a fusion feature vector, which contains both sequence features reflecting recent dynamic changes and index correlation features reflecting long-term correlation patterns;

[0150] The dimension of the fusion feature vector is 2d h , which contains both recent dynamic and long-term correlation information, and the calculation formula of the fusion feature vector h fuse is:

[0151]

[0152] In the formula, h seq is the sequence dynamic feature vector; is a vector splicing operator;

[0153] S504. The fusion feature vector is input into a fully connected layer and processed by a Softmax activation function to output a probability distribution vector of the fault class of each sample, and each element of the probability distribution vector represents the predicted probability that the sample belongs to the corresponding fault class;

[0154] The dimension of the fault class probability distribution vector of the i th sample is C, and the c th element of the fault class probability distribution vector of the i th sample is , which represents the probability that the i th sample belongs to the c th class, c is a class index, c ∈ {1, 2, …, C}, C is the total number of fault classes, and the calculation formula of the fault class probability distribution vector of the i th sample is:

[0155]

[0156] In the formula, Wo is the output layer weight matrix with dimension C x 2d h is a learnable parameter; b o is the output layer bias vector with dimension C, which is a learnable parameter.

[0157] Parallel extraction of sequence dynamic feature vectors h seq and index associated feature vectors h core and fusion, effectively taking into account the differences in time and space scales, and thus improving the detection accuracy of multiple types of faults.

[0158] Internet data center fault detection needs to capture both recent mutation features such as CPU abnormal surge and long-term associated patterns such as memory leakage trend. Single path feature extraction cannot take into account the differences in time and space scales. Conventional single-channel fault detection models cannot effectively fuse multi-scale features, resulting in missed detection or false positives. S5 extracts sequence dynamic features and index associated features in parallel, and uses a dual-channel fusion mechanism to integrate the advantages of the two types of features.

[0159] S6. Calculate the dynamic category weight coefficient of each category, and calculate the adaptive loss function.

[0160] S601. For each fault category, calculate the dynamic weight coefficient of each category according to the frequency of each category in the training set and the human preset cost factor. The dynamic weight coefficient of each category fuses category distribution information and business security requirements. The greater the value, the higher the importance of the category in the loss calculation, and the influence strength of the category frequency on the weight is controlled through the frequency sensitive coefficient in the calculation process.

[0161] The greater the value of the dynamic weight coefficient of the cth category, the higher the importance of the category in the loss function. The dynamic weight coefficient of the cth category γ c is calculated as follows:

[0162]

[0163] In the formula, β is the frequency sensitive coefficient, which controls the strength of the frequency influence. When β→0, the frequency influence is ignored. When β→+∞, the frequency dominates the weight. β can be set to 1.0. freq(c) is the frequency of the cth category, which represents the frequency of the cth category in the training set. The calculation method is represented as N c / N, N c is the number of samples of the cth category; c ′ is the category index different from c, c∈{1,2,…,C}; freq(c ′ ) is the frequency of the cth ′ category, which represents the frequency of the cth c′ category in the training set. The calculation method is represented as N c′ / N, N′ The number of samples in each category; η c The cost factor for the c-th category is preset by humans, and can be set as η1 = 2.0, η2 = 0.5;

[0164] Category cost factor η c Based on historical fault data or expert knowledge, we can set c=1 to correspond to the disk fault category, where the cost of misjudging it as a normal category is high, so we can set η1=2.0. c=2 corresponds to the temperature fluctuation category, where the cost of misjudging it is lower, so we can set η2=0.5.

[0165] S602. Combining the calculated dynamic weight coefficients of each category and the focus loss mechanism, an adaptive loss function is constructed. The adaptive loss function, based on the standard cross-entropy loss, weights the loss term for each sample and each category. On the one hand, the dynamic weight coefficients are used to scale the loss of different categories, making the fault detection model pay more attention to key fault categories with few samples or high misclassification costs. On the other hand, combined with the focus loss term, the loss contribution of easily classified samples with predicted probabilities close to the true labels is reduced, making the fault detection model more focused on learning difficult samples.

[0166] The formula for calculating the adaptive loss function Loss is:

[0167]

[0168] In the formula, To predict the probability that the i-th sample belongs to the c-th class in the fault detection model; α is the focusing parameter, which reduces the loss contribution of easily distinguishable samples, and can be set to α = 2; I(·) is the indicator function, where I(y i =c) In item y, when y i Output 1 if = c, otherwise output 0; y i Let be the true class label of the i-th sample; log(·) is the logarithmic function, with the default base being the natural constant.

[0169] Through dynamic weighting coefficient γ c The dynamic fusion category frequency freq(c) and the preset cost factor η c This allows the loss function to adapt to both uneven data distribution and business security requirements, thereby improving the detection rate of high-risk faults.

[0170] The internet data center failure categories are seriously unbalanced, and the misclassification costs of different failure types are significantly different. For example, misclassifying "disk failure" as "normal" is more serious than the opposite case. Standard cross-entropy loss treats all categories equally and cannot solve the problems of category imbalance and cost sensitivity at the same time, resulting in insufficient identification of key failure types by the failure detection model. S6 adopts an adaptive loss function that dynamically adjusts the category weight coefficient while considering the category frequency and the preset misclassification cost.

[0171] S7. Calculate the hierarchical gradient, perform the failure sensitivity oriented gradient clipping, perform the hierarchical parameter update and non-negative projection, and obtain the trained failure detection model;

[0172] S701. Calculate the total loss function value for the current batch of samples, and calculate the gradients of the gated recurrent unit parameter set, the non-negative Tucker decomposition parameter set, and the fully connected classification layer parameter set, respectively.

[0173] Gated recurrent unit parameter gradient The calculation formula is:

[0174]

[0175] In the formula, is the partial derivative symbol; θ GRU is the gated recurrent unit parameter set, including the reset gate weight matrix W r , the reset gate bias vector b r , the feature transformation weight matrix W c , the attention weight matrix W k , the attention bias vector b c , the update gate weight matrix W z , the candidate state weight matrix W h , the update gate bias vector b z , and the candidate state bias vector b h .

[0176] Non-negative Tucker decomposition parameter gradient The calculation formula is:

[0177]

[0178] In the formula, θ Tucker is the non-negative Tucker decomposition parameter set, including the first modal factor matrix U (1) , the second modal factor matrix U (2) , and the third modal factor matrix U (3) .

[0179] Fully connected classification layer parameter gradient The calculation formula is:

[0180]

[0181] where θ FC is the set of fully connected classification layer parameters, including the output layer weight matrix W o , the output layer bias vector b o .

[0182] Except for the trainable parameters in the set of gated recurrent unit parameters, the set of non-negative Tucker decomposition parameters and the set of fully connected classification layer parameters, the conventional gradient descent method is used for separate training.

[0183] S702. Calculate the average sampling probability of the current batch, and dynamically adjust the threshold of gradient clipping using the average sampling probability. A batch with a high average sampling probability means that it contains more high-risk samples, and will obtain a larger clipping threshold. The gradient information of the key batch samples can greatly affect the parameter update direction;

[0184] The calculation formula of the gradient clipping threshold τ cs is as follows:

[0185]

[0186] where τ base is the basic clipping threshold, which can be set as τ base = 0.1; λ clip is the clipping intensity coefficient, which can be set as λ clip = 0.3; P batch is the batch sampling probability; μ P is the mean of the training set sampling probability, which is pre-calculated from the training set and frozen; σ P is the standard deviation of the training set sampling probability, which is pre-calculated from the training set and frozen;

[0187] The calculation formula of the batch sampling probability P batch is as follows:

[0188]

[0189] where B is the batch size, i.e. the number of samples in each training batch; P sample (X i ) is the sampling probability of the i-th sample;

[0190] The calculation formula of the sampling probability P sample (X i ) of the i-th sample is as follows:

[0191]

[0192] where λ wFor the failure sensitivity coefficient, λ w = 0.2; F(X i ) is the failure sensitivity score of the i-th sample; N is the total number of samples in the training set; F(X j ) is the failure sensitivity score of the j-th sample;

[0193] S703. The three types of parameter gradients of the gated recurrent unit parameters, the non-negative Tucker decomposition parameters and the fully connected classification layer parameters are respectively clipped using a gradient clipping threshold, so that the gradient values are limited within a certain range, and the specific clipping method is as follows:

[0194] Based on the gradient clipping threshold τ cs , each element value in the gradient of the gated recurrent unit parameters the gradient of the non-negative Tucker decomposition parameters and the gradient of the fully connected classification layer parameters is limited, that is, all numerical values in the gradient tensor or vector are traversed, if the numerical value is greater than τ cs , it is set to τ cs , if the numerical value is less than -τ cs , it is set to -τ cs , and if the numerical value is within the interval [-τ cs , τ cs ], it remains unchanged, so that all gradient values are strictly constrained within the range [-τ cs , τ cs ], so as to control the gradient amplitude and enhance the stability of the optimization process.

[0195] Different learning rates are set for the gated recurrent unit parameters, the non-negative Tucker decomposition parameters and the fully connected classification layer parameters respectively, and the clipped gradients and the corresponding learning rates are used to update the three types of parameters respectively.

[0196] The updated set of gated recurrent unit parameters θ GRU is represented as:

[0197]

[0198] In the formula, ← is a parameter update operation; η GRU is the learning rate of the gated recurrent unit, which can be set as η GRU = 0.001; is the clipped gradient of the gated recurrent unit parameters;

[0199] The updated set of non-negative Tucker decomposition parameters θ Tucker is represented as:

[0200]

[0201] In the formula, ηTucker For the learning rate of non-negative Tucker decomposition, it can be set is the clipped non-negative Tucker decomposition parameter gradient;

[0202] The parameter set of the fully connected classification layer is θ FC The calculation formula is:

[0203]

[0204] In the formula, η FC For the learning rate of the fully connected classification layer, it can be set η FC = 0.005; is the clipped fully connected classification layer parameter gradient;

[0205] The internet data center fault detection model contains various types of trainable parameters, and the convergence characteristics and sensitivity to faults in training are significantly different. A unified learning rate mechanism cannot adapt to such differences, which may lead to insufficient or unstable updates of parameters related to key fault features. S7 adopts a hierarchical adaptive matrix estimation optimizer. Non-negative constraints are imposed on the updated non-negative Tucker decomposition parameters, and negative values are set to zero by traversing all factor matrix elements to ensure that the decomposition results strictly satisfy the non-negative constraints and maintain the physical interpretability of the features, such as the correlation strength of CPU load, temperature, and other indicators, which cannot be negative.

[0206] Through S2, the data adaptation construction of the input layer of the fault detection model is completed, through S3, the associated feature modeling of the fault detection model is completed, through S4, the key feature strengthening of the fault detection model is completed, through S5, the multi-scale feature fusion and classifier construction of the feature layer and the output layer of the fault detection model are completed, through S6, the loss mechanism design of the optimization layer of the fault detection model is completed, through S7, the parameter optimization mechanism construction of the optimization layer of the fault detection model is completed, and the fault detection model is constructed based on the gated recurrent unit. The complete logic chain is: input layer optimization → associated feature modeling → feature optimization → multi-scale feature fusion and classification → optimization target design → parameter training mechanism.

[0207] S8. Utilize the trained fault detection model to automatically diagnose new real-time monitoring data;

[0208] S801. Collect new monitoring data, divide it into samples according to the time window, and input it into the trained fault detection model;

[0209] S802. Perform S2, S3, S4, S5 in turn to obtain a state category probability distribution, and select the category with the highest probability as the fault diagnosis result; when the output category is a disk failure, a memory leak, a network interruption, a CPU overload, or a temperature anomaly, trigger an alarm notification of the corresponding level, and mark the fault occurrence period and the associated index anomaly pattern to provide precise fault positioning basis for operation and maintenance personnel.

[0210] (1) To verify the effectiveness of the dynamic importance sampling strategy in improving fault detection performance, a dynamic importance sampling mechanism analysis experiment was conducted. This experiment used a monitoring data set of a large Internet data center for 3 consecutive months, containing about 25,000 time series records, covering 5 index dimensions of CPU load, memory usage, network traffic, disk I / O, and temperature. The data set was divided into samples by 10-minute time windows, generating about 45,000 samples, of which normal samples accounted for 92.6%, and fault samples accounted for 7.4%. Fault categories included disk failure (1.2%), memory leak (1.8%), network interruption (2.1%), CPU overload (1.5%), and temperature anomaly (0.8%);

[0211] The comparison methods in this experiment include:

[0212] 1) Random sampling: the probability of each sample being sampled is equal; 2) SMOTE (Synthetic Minority Over-sampling Technique): synthesize minority class samples by linear interpolation, set the nearest neighbor parameter to 5; 3) ADASYN (Adaptive Synthetic Sampling Method for Imbalanced Learning): synthesize minority class samples adaptively according to sample density, set the nearest neighbor parameter to 5; 4) Dynamic importance sampling provided by the present application;

[0213] In this experiment, the training set accounted for 70%, and the test set accounted for 30%. All methods used the same division, and the evaluation index used macro-averaged F1-score to take into account the balance of all categories. The experiment was repeated 10 times to take the average value and eliminate the influence of randomness.

[0214] (1) Compare the detection performance of each method on each fault. Use the same GRU classifier architecture and process the training set with four methods respectively. Train each method 10 times and take the average F1 value. When testing, use a unified test set (containing all fault types) to evaluate the F1 score of each fault, and get Figure 1 . As shown in Figure 1 , dynamic importance sampling achieves the highest detection performance on all fault categories, with a more significant advantage in the disk failure and temperature anomaly categories with few fault samples, indicating that dynamic sampling probability calculation successfully focuses on high fault risk samples, effectively alleviating the class imbalance problem.

[0215] (2) Analyze the sampling probability difference between normal samples and fault samples, randomly select 1000 normal samples and 1000 fault samples from the training set, cover all fault types, calculate the sampling probability of each sample, and draw the probability distribution curve using kernel density estimation, and get Figure 2 , Figure 2 Intuitively present the sampling probability difference between normal samples and fault samples, and the fault samples show right skewed distribution, that is, the high probability area is dense, which proves the precise recognition ability of the fault sensitivity score function to abnormal samples.

[0216] (B) To evaluate the ability of non-negative Tucker decomposition in extracting cross-index dynamic interaction features, and verify its advantages compared with traditional dimension reduction methods, a multi-index correlation tensor decomposition effect analysis experiment is carried out;

[0217] The same data set as experiment (I) is used in this experiment, and 10000 samples are randomly selected for feature extraction comparison in this experiment; The comparison methods of this experiment include: 1) PCA (principal component analysis): retaining 95% of the principal components of the variance; 2) LDA (linear discriminant analysis): linear projection that maximizes the class distinction; 3) t-SNE (t-distribution stochastic neighbor embedding): the number of iterations is set to 1000 times; 4) Autoencoder (autoencoder): adopts a three-layer structure, the encoding dimension is 64-32-16, and the activation function adopts ReLU activation;

[0218] 5) The tensor decomposition provided by the present application;

[0219] After feature extraction, the same classifier is connected for performance comparison, specifically using a two-layer perceptron network as the classifier.

[0220] (1) Analyze the correlation pattern between indexes extracted by tensor decomposition, calculate the correlation matrix of the factor matrix obtained by non-negative Tucker decomposition, and draw a heat map after row and column normalization to get Figure 3 , Figure 3 Show the correlation pattern between indexes extracted by tensor decomposition, where the combination of key indexes such as CPU load, disk I / O, memory usage and temperature shows high intensity correlation (warm color area), which proves that the method successfully captures the physical coupling relationship of device operation;

[0221] (3) By reducing the feature dimension to 2D, the core tensor feature distribution is visualized to get Figure 4 , specifically, the core tensor of the test set sample is flattened into a vector, and t-SNE (confusion degree = 30, learning rate = 200) is used to reduce it to 2D, and normal and fault samples are labeled, as shown in Figure 4 The normal samples (blue) and fault samples (orange) form obvious separated clusters in the core feature 1-2 plane and the nonlinear decision boundary verifies the strong discriminability of the feature;

[0222] (Three) To verify the effectiveness of the adaptive weighting mechanism of index importance and the synergistic gain of the dual-channel feature fusion strategy, an adaptive weighting and dual-channel fusion effect analysis experiment is conducted;

[0223] The same data set as in Experiment (One) is used in this experiment, and 1,200 samples containing obvious fault evolution are selected from the data set, with 200 samples for each fault category. The changes in indicators within 1 hour before the occurrence of the fault are analyzed. The comparison methods in this experiment include: 1) only sequence features: only using GRU to extract time series dynamic features; 2) only correlation features: only using core tensor to extract index correlation features; 3) feature concatenation: direct concatenation without adaptive weighting; 4) dual-channel fusion of the present technology. The evaluation indicators set in this experiment include precision, recall, F1-score, and detection delay (the time from fault occurrence to alarm).

[0224] (1) By analyzing the dynamic changes of index importance weights over time, we get Figure 5 , select a real disk failure event, calculate the weight of each indicator at 1 minute intervals, draw the weight change curve, Figure 5 reveal the dynamic evolution law of the index weight, the moment of fault occurrence (black dotted line) the weight of disk I / O increases sharply, the weight of network traffic rises instantaneously, verifying the real-time focusing ability of the learnable weight generator on key indicators;

[0225] (2) The experiment compares the performance differences of sequence features, correlation features, feature concatenation, and dual-channel fusion of four strategies. The same classifier is used for the four strategies, specifically a two-layer perceptron as the classifier. The macro F1, precision, recall, AUC, and robustness (performance variance inverse under different loads) are evaluated on the test set, and we get Figure 6 , Figure 6 showing that dual-channel fusion leads in precision, recall, and comprehensive evaluation indicators, proving the complementary advantages of sequence dynamic features and index correlation features;

[0226] (3) Visualize the dual-channel feature space distribution, reduce the fusion features to 2D through PCA, and draw the sample scatter plot to get Figure 7 , as shown in Figure 7 , dual-channel features make fault samples form a tight cluster and are far away from normal samples, and the decision boundary presents a complex nonlinear form, which explains the reason for the improvement of classification performance from a geometric perspective;

[0227] (Four) To evaluate the comprehensive performance and resource efficiency of the present technology in real deployment environment, the overall system performance and resource efficiency analysis experiment is conducted;

[0228] The same data set as experiment (1) is used, and 12,000 samples (normal samples account for 88.7%) covering 5 fault types are selected. The comparison methods of this experiment include: 1) standard GRU (standard gated recurrent unit): the number of hidden units is 128, and the structure is double-layer; 2) LSTM-AE (long short-term memory-autoencoder): encoder-decoder structure, and the hidden dimension is 64; 3) CNN+Attention (convolutional neural network combined with attention mechanism): the convolution kernel size is [3, 5, 7], and the number of attention heads is 4; 4) Transformer: the number of encoder layers is 3, the number of attention heads is 8, and the hidden dimension is 512; 5) the complete model of the present technology;

[0229] (1) The F1 score, precision, and recall of the five types of models on the unified test set are compared to obtain Figure 8 , Figure 8 which shows that the present technology leads in all performance indicators, especially in the recall rate (fault recall rate), proving the synergistic effect of the complete method system;

[0230] (2) The false alarm rate trend over time is analyzed. The false alarm rate (false alarm number / total alarm number) is calculated every 12 hours for 30 consecutive days, and the curves of the present technology and the best benchmark (LSTM-AE) are drawn to obtain Figure 9 As shown in Figure 9 , the curve of the present technology always runs smoothly at a low level, and the fluctuation range (orange translucent area) is much smaller than that of the benchmark method, which reflects the continuous inhibition ability of the adaptive loss function to false alarms;

[0231] (3) The performance under different load conditions is analyzed. By adjusting the system load during data collection, three load conditions are simulated, and the F1 score of each model on the three load test sets is evaluated to obtain Figure 10 As shown in Figure 10 , as the load increases, the performance of the present technology decays the least, and it still maintains high detection ability at the peak load, which highlights the robustness of the algorithm.

Claims

1. A method for fault detection in Internet data centers based on artificial intelligence, characterized in that, include: S1. Collect monitoring data from Internet data centers and label the monitoring data with status; S2. Calculate the dynamic sampling probability of each sample based on the fault sensitivity score of each sample; S3. Construct the spatiotemporal correlation tensor of each sample, and decompose the constructed spatiotemporal correlation tensor of each sample into the product of the core tensor and multiple factor matrices; S4. Generate a weighted time series data matrix for each sample based on the normalized importance weights of each dimension of the indicators; S5. Extract sequence dynamic features, extract indicator correlation features, perform dual-channel feature fusion, and generate fault classification probability; S6. Calculate the dynamic category weight coefficients for each category, and calculate the adaptive loss function; S7. Calculate the hierarchical gradient, perform fault sensitivity-guided gradient pruning, and perform hierarchical parameter updates and non-negative projection; S8. Automated fault diagnosis is performed on newly acquired real-time monitoring data using the trained fault detection model.

2. The method for fault detection in an internet data center based on artificial intelligence according to claim 1, characterized in that, The steps in S1 include: S101. Deploy multiple types of sensors in key areas of the Internet data center to continuously collect monitoring data on key performance indicators and environmental parameters that reflect the system's operating status; S102. Label the time period of each sample according to its running status.

3. The method for fault detection in an internet data center based on artificial intelligence according to claim 1, characterized in that, The steps in S2 include: S201. Calculate the absolute deviation of all indicators for each sample at all time points that exceed three times the standard deviation threshold of the indicator. Sum all the obtained absolute deviation values ​​to obtain the fault sensitivity score of each sample. S202. Based on the calculated fault sensitivity scores of all samples, normalization is performed using the Softmax function to calculate the sampling probability of each sample.

4. The method for fault detection in an internet data center based on artificial intelligence according to claim 1, characterized in that, The steps in S3 include: S301. Perform an outer product operation on the original temporal data matrix of each sample to generate a three-dimensional spatiotemporal correlation tensor for each sample; S302. Decompose the spatiotemporal correlation tensor of each sample into a product of the core tensor and multiple factor matrices.

5. The method for fault detection in an Internet data center based on artificial intelligence according to claim 1, characterized in that, The steps in S4 include: S401. Flatten the core tensor of each sample into a one-dimensional vector and input it into the multilayer perceptron network. The multilayer perceptron network outputs the original importance score vector of each sample. S402. The original importance score vector of each sample is converted into normalized importance weights of each dimension index of each sample by using the Softmax function with temperature coefficient. S403. Multiply the normalized importance weights of each dimension of each sample element-wise with the original time series data matrix to generate a weighted time series data matrix for each sample.

6. The method for fault detection in an Internet data center based on artificial intelligence according to claim 1, characterized in that, The steps in S5 include: S501. Input the weighted time series data matrix of each sample into the gated recurrent unit network based on association perception, process the entire time series, and take the hidden state vector output by the last time step of the network as the sequence dynamic feature vector; S502. The core tensor obtained through non-negative Tucker decomposition is flattened into a one-dimensional vector and input into a multilayer perceptron network. It is then mapped to a vector of the same dimension as the sequence dynamic feature vector to obtain the index-related feature vector. S503. The extracted sequence dynamic feature vector and indicator-related feature vector are concatenated to generate a fused feature vector; S504. Input the fused feature vector into the fully connected layer, process it through the Softmax activation function, and output the probability distribution vector of the fault category of each sample.

7. The method for fault detection in an internet data center based on artificial intelligence according to claim 1, characterized in that, Step S6 includes: S601. For each fault category, calculate the dynamic weight coefficient of each category based on the frequency of each category in the training set and the cost factor preset by the user; S602. Combine the calculated dynamic weight coefficients for each category with the focus loss mechanism to construct an adaptive loss function.

8. The method for fault detection in an Internet data center based on artificial intelligence according to claim 1, characterized in that, The steps in S7 include: S701. For the total loss function value calculated for the current batch of samples, calculate its gradient with respect to the gated recurrent unit parameter set, the non-negative Tucker decomposition parameter set, and the fully connected classification layer parameter set, respectively; S702. Calculate the average sampling probability of the current batch and dynamically adjust the gradient clipping threshold using the average sampling probability; S703. Use gradient clipping thresholds to clip the gradients of the three types of parameters: gated recurrent unit parameters, non-negative Tucker decomposition parameters, and fully connected classification layer parameters, respectively, limiting the gradient values ​​to a specific range.

Images