An edge cloud failure prediction method and system based on deep learning

Abstract

The application discloses an edge cloud fault prediction method and system based on deep learning, comprising: collecting edge cloud equipment fault information and constructing fault feature samples; using an unsupervised learning method to train non-fault data in the fault feature samples, obtaining a CPU fault self-encoding model, an equipment fault self-encoding model, a memory fault self-encoding model and a network fault self-encoding model, and using mean absolute error to calculate the probability of CPU fault, equipment fault, memory fault and network fault of the samples; sequencing all fault probabilities, and performing feature engineering processing on the fault feature samples to obtain a fault feature vector, using an LSTM network model, a DeepAR model and a linear classifier to construct a fault time series prediction model for predicting the fault vector; and using an FM model to train the fault vector to obtain whether the equipment has a fault. The application can further mine the features of high-risk equipment faults and solve the problem of low recall rate.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault prediction, and particularly relates to an edge cloud fault prediction method and system based on deep learning, which can be applied to an edge cloud scene and reused in multiple scenes such as a central cloud, an IDC, a CDN, and a terminal device. BACKGROUND

[0002] With the advent of the Internet of Everything era and the popularization of 5G communication networks, the scale of Internet data is growing exponentially. Under this background, the centralized architecture of traditional cloud computing has been unable to meet the needs of terminal users for timeliness, capacity, and computing power. The ultra-low latency, massive data, and edge intelligence of edge cloud have prompted more enterprises to choose edge cloud technology solutions, and have made edge cloud computing an important component built between the central cloud and the terminal in the market. Compared with the central cloud, the edge cloud not only has a complex network structure of the Internet, Ethernet, 5G, WIFI, etc., but also has a large number of hardware nodes with various hardware configurations. This complex edge cloud structure results in not only a higher failure rate of edge cloud devices than that of the central cloud, but also more external factors causing failures. In the operation and management deployment of the edge cloud, accurate determination and prediction of the failure rate of devices are not only crucial to the service stability of the edge cloud, but also very important to the market profitability of edge cloud service providers.

[0003] Existing fault prediction technologies generally use supervised classification or time series prediction and other related prediction technologies, which cannot solve the actual scenarios of complex edge cloud environments and variable businesses, and have the following shortcomings: the scenarios that can be predicted are relatively single, and cannot adapt to various complex and variable scenarios such as the region of edge cloud device deployment, the business of edge cloud device deployment, population, and space-time; the recall rate of the model for faulty devices is not high, because the probability of hardware device failure is generally low in most scenarios, so there is a very obvious positive and negative imbalance phenomenon in the actual collected data, which also leads to the fact that most models cannot truly predict device failures; the optimization target is relatively single, because the classification model is usually divided into multiple models for training or transformed into multi-classification for training when predicting the failure probability of different hardware devices, and this approach cannot consider the correlation between optimization targets; it cannot solve the cascading effect caused by part of the risk explosion, and a simple classification model is not sensitive to the cascading effect or does not have a quantitative standard for the consequences of part of the burst information. SUMMARY

[0004] To solve the above technical problems, the technical scheme adopted by the present application is as follows:

[0005] A deep learning-based edge cloud fault prediction method, comprising the following steps:

[0006] S1, collect the fault information of the edge cloud device, and construct a fault feature sample;

[0007] S2, use an unsupervised learning method to expand the training of non-fault data in the fault feature sample obtained in step S1, obtain a CPU fault auto-encoding model, a device fault auto-encoding model, a memory fault auto-encoding model, and a network fault auto-encoding model, and use the mean absolute error to calculate the probability of each sample in the fault feature sample to produce CPU fault, device fault, memory fault and network fault;

[0008] S3, serialize the fault probabilities output by all fault auto-encoding models, and perform feature engineering processing on the fault feature sample in step S1 to obtain a fault feature vector F, and use an LSTM network model, a DeepAR model and a linear classifier to construct a fault time series prediction model for predicting the fault vector;

[0009] S4, use the FM model to determine and train the fault vector obtained in step S3 to obtain whether the device will fail.

[0010] In step S1, the fault feature sample includes device ID, service ID, device feature, service feature, fault date, CPU fault, memory fault, network fault, and device fault.

[0011] The step S2 comprises the following steps:

[0012] S2.1, mark the data without CPU fault as positive samples from the fault feature sample, and filter the positive samples using a multivariate Gaussian distribution;

[0013] S2.2, use the auto-encoding model to train and optimize the filtered non-CPU fault data to obtain a CPU fault auto-encoding model for predicting the CPU fault probability;

[0014] S2.3, inputting the fault feature sample of step S1 into the CPU fault auto-encoding model for testing, calculating the mean absolute error according to the vector length of the input non-CPU fault data and the vector length output by the model, and normalizing all the mean absolute errors to obtain the probability of the fault feature sample generating the CPU fault;

[0015] S2.4, constructing the device fault auto-encoding model, the memory fault auto-encoding model and the network fault auto-encoding model according to the method of steps S2.1-S2.3, and calculating the probability of each sample in the fault feature sample generating the device fault, the probability of generating the memory fault and the probability of generating the network fault.

[0016] In step S2.1, the filtering of the positive samples by using the multivariate Gaussian distribution refers to filtering out abnormal data with a density probability p(x) less than a threshold value ξ, where the density probability p(x) is calculated according to the following formula:

[0017]

[0018] In the formula, μ represents the mean, m represents the number of samples, x represents the sample, and ∑ represents the covariance.

[0019] In step S2.2, the hidden layer of the auto-encoding model is three layers, and the number of nodes of the three-layer network is <64, 32, 64>.

[0020] The step S3 includes the following steps:

[0021] S3.1, serializing all the fault probabilities obtained in step S2 to obtain a CPU fault probability sequence, a memory fault probability sequence, a network fault probability sequence and a device fault probability sequence, and splicing all the fault probability sequences with historical fault probabilities to obtain a full-amount fault probability sequence;

[0022] S3.2, performing feature engineering processing on the fault feature sample in step S1 to obtain a fault feature vector F, and updating the full-amount fault probability sequence by using the fault feature vector F and the fault feature sample;

[0023] S3.3, respectively cutting the updated full-amount fault probability sequence according to the fault type, inputting the cut sequence data into an LSTM model for training to obtain an LSTM_CPU model, an LSTM_memory model and an LSTM_network model, and inputting the updated full-amount fault probability sequence into a DeepAR model for optimization training to obtain a fault distribution prediction model;

[0024] S3.4, splicing all the models obtained in step S3.3 to obtain a fault timing prediction model, and training a fault vector output by the fault timing prediction model using a linear classifier to obtain the weight of each dimension vector in the fault vector.

[0025] In step S3.3, the model structure of the LSTM_CPU model, the LSTM_memory model and the LSTM_network model all includes a first LSTM, a second LSTM, a third LSTM, a Dropout layer and a full connection layer connected in sequence.

[0026] In step S4, the determination formula is:

[0027]

[0028] In the formula, y represents a fault classification target value, w0 represents a bias, w i represents the weight of the vector at the i-th position in the model input vector, x i represents the vector value at the i-th position in the model input vector, n represents the length of the model input vector, v if represents the vector value at the j-th position in the hidden vector of x i , and k represents the length of the preset hidden vector.

[0029] An edge cloud fault prediction system based on deep learning, comprising

[0030] A sample collection module is configured to collect fault information of the edge cloud device, and construct a fault feature sample according to the fault information.

[0031] A fault auto-encoding model establishment module is configured to train non-fault data in the fault feature sample based on an unsupervised learning method, and construct a CPU fault auto-encoding model, a device fault auto-encoding model, a memory fault auto-encoding model and a network fault auto-encoding model.

[0032] A fault probability calculation module is configured to calculate a mean absolute error according to a corresponding input vector of the fault feature sample and output vectors of the CPU fault auto-encoding model, the device fault auto-encoding model, the memory fault auto-encoding model and the network fault auto-encoding model, and obtain a probability of CPU fault, a probability of device fault, a probability of memory fault and a probability of network fault according to the mean absolute error.

[0033] The fault time sequence prediction model construction module is used for performing feature engineering processing on the fault feature samples to obtain fault feature vectors, and is used for constructing a fault time sequence prediction model for predicting the fault vectors based on the serialized CPU fault probability, the device fault probability, the memory fault probability and the network fault probability, and using an LSTM network model, a DeepAR model and a linear classifier.

[0034] The present application has the following beneficial effects:

[0035] The present application fully utilizes various multi-dimension feature data to consider various scene factors when predicting faults, and converts a classification problem with very few fault samples into a probability prediction problem with numerous samples through a self-encoder. In the process of converting classification into probability, the influence of various scene factors is fully considered, the characteristics of high-risk device faults are better mined, and the problem of low recall rate of fault devices is solved. Through a multi-objective prediction algorithm, the prediction target weight and the correlation of each hardware fault in the device can be better measured. Through a combination of a self-encoder, time sequence prediction and FM, not only can the level connection effect caused by risk explosion be solved, but also the uncertain factors caused by space-time changes can be better solved. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description.

[0037] Figure 1 The present application is a flowchart.

[0038] Figure 2 The present application is a structure diagram of a fault time sequence prediction model.

[0039] Figure 3 The present application is a training flowchart of an FM model.

[0040] Figure 4 The present application is a training effect diagram of an LSTM_CPU model. DETAILED DESCRIPTION

[0041] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort belong to the scope of the present application.

[0042] The auto-encoder model, also known as the Auto-Encoder model, is a neural network algorithm based on the back propagation algorithm and the optimization method, which is divided into an encoder (Encoder) and a decoder (Decoder). DeepAR is a prediction algorithm proposed by Amazon for unified modeling of a large number of related time series. The algorithm uses deep learning technology to train an autoregressive recurrent network model on a large number of time series, which can effectively learn a global model from related time series and learn complex patterns, such as seasonality and increasing uncertainty of data over time, to predict each time series. Long Short-Term Memory (LSTM) is a time recurrent neural network designed to solve the long-term dependence problem of general recurrent neural networks (RNN). All RNNs have a chain form of repeated neural network modules. In a standard RNN, this repeated structure module has only a very simple structure, such as a tanh layer. FM, the FM (Factorization Machine) algorithm can be used for regression and binary classification prediction. It considers the interaction between features and is a nonlinear model.

[0043] Embodiment 1: A deep learning-based edge cloud fault prediction method, as shown in Figure 1 includes the following steps:

[0044] S1, collect the fault information of the edge cloud device, and construct a fault feature sample, including the following steps:

[0045] S1.1, collect the device basic data, the service basic data running on the device, the device fault data and the historical monitoring data of each edge cloud device;

[0046] The device basic data of the edge cloud hardware includes device ID and device feature data, and the device feature data includes CPU model, CPU number, CPU factory date, memory size, disk size, hdd disk number, ssd disk number, hdd disk average disk iops, ssd disk average iops, device location, device city, device operator (mobile, Unicom or Telecom, etc.).

[0047] The business basis data includes a business ID and business characteristic data, and the business characteristic data includes a previous day traffic of the business, a number of business deployment nodes, a previous day average traffic of the business deployment nodes, and the like.

[0048] The device fault data includes a fault date, whether the CPU is faulty, whether the memory is faulty, whether the network is faulty, whether the device is faulty, a corresponding business ID and device ID when the fault occurs, and the like.

[0049] The historical monitoring data includes a device ID, a 95% CPU usage rate, a 95% memory usage rate, a 95% bandwidth usage rate, and a network disconnection rate. All collected data are collected through a database, log4j, and the like.

[0050] S1.2, constructing a fault state sample using the data obtained in step S1.1;

[0051] The fault state sample includes a device ID, a business ID, a fault date, whether the CPU is faulty, whether the memory is faulty, whether the network is faulty, and whether the device is faulty. For example, a fault state sample example is {device_id='dfweropjp', buss_id='kuaishou', dt='20220101', is_cpu_bad=0, is_memory_bad=0, is_band_bad=0, is_bad=0}, which indicates that the device number 'dfweropjp', the business ID 'kuaishou', the date '20220101', whether the CPU is faulty '0', whether the memory is faulty '0', whether the network is faulty '0', and whether the device is faulty '0'.

[0052] S1.3, associating the fault state sample constructed in step S1.2 with the data in step S1.1 to form a fault characteristic sample;

[0053] The fault feature sample includes device ID, service ID, device feature, service feature, fault date, whether CPU is faulty, whether memory is faulty, whether network is faulty, and whether device is faulty. For example, { 'device ID' = dfweropjp,'service ID' = 'kuaishou', 'device feature' = [device id, cpu model, cpu number, cpu factory date, memory size, disk size, disk number, hdd disk number, ssd disk number, hdd disk average disk iops, ssd disk average iops, province where the device is located, city where the device is located, device operator],'service feature' = [service id, service day before traffic, service deployment node number, service deployment node average traffic before], 'fault date' = '20220101', 'whether CPU is faulty' = 0, 'whether memory is faulty' = 0, 'whether network is faulty' = 0, 'whether device is faulty' = 0}.

[0054] S2, using unsupervised learning method to expand training non-fault data in the fault feature sample obtained in step S1, obtaining CPU fault auto-encoding model, device fault auto-encoding model, memory fault auto-encoding model and network fault auto-encoding model, using mean absolute error to calculate the probability of each sample in the fault feature sample to produce CPU fault, device fault, memory fault and network fault, including the following steps:

[0055] S2.1, the data without CPU fault is selected from the fault feature sample and marked as positive sample, and the positive sample is filtered by using multivariate Gaussian distribution;

[0056] The data without CPU fault is filtered out, that is, the data equal to 0 in the 'whether CPU is faulty' field is filtered out, and then the abnormal value is filtered by using multivariate Gaussian distribution. Specifically, the density probability p(x) of each sample in the tenth number is set as the threshold value ξ, and the abnormal data with p(x) < ξ is filtered out. By filtering out the abnormal data, the accuracy of the auto-encoder model in S2.2 can be further improved, so that the root mean square error (RMSE) is reduced from 0.043 to about 0.039, and the mean squared error (MSE) index of the auto-encoding is improved by 1.5%.

[0057] The formula of the parameters of the multivariate Gaussian distribution is:

[0058]

[0059]

[0060]

[0061] wherein μ represents a mean value, m represents a sample number, x represents a kth sample, ∑ represents a covariance, and p(x) represents a density probability of the sample x. k

[0062] S2.2, training the filtered non-CPU fault data by using the auto-encoding model to obtain a CPU fault auto-encoding model for predicting a CPU fault probability;

[0063] The fault feature samples include CPU model, province where the device is located, city where the device is located, device operator, and service ID, which are classified data, and other data are non-classified data. First, the classified data in the filtered non-CPU fault data obtained in step S2.1 is one-hot coded, then the one-hot coded classified data and the non-classified data in the filtered non-CPU fault data are spliced, so that the input data is converted into a vector with a length of 138, and then the vector is trained.

[0064] The hidden layer of the auto-encoding model is three layers, the node numbers of the three layers are <64, 32, 64> respectively, the last layer is an output with 138 nodes, and the mean square error is used as the loss function of the auto-encoding model, and the final output of the model is a vector with a length of 138. Through experiments, it is verified that the average absolute error (MAE) index is improved by 0.07% compared with other node configurations.

[0065] S2.3, inputting the fault feature samples of step S1 into the CPU fault auto-encoding model for testing, calculating the average absolute error according to the vector length of the input non-CPU fault data and the vector length of the model output, and normalizing all the average absolute errors to obtain the probability of the fault feature samples generating CPU fault;

[0066] The fault feature samples are full data, that is, they include fault data and non-fault data, the CPU fault auto-encoding model outputs a vector with a length of 138, the MAE of the 138-length vectors of the input and the output is calculated, the higher the MAE, the higher the CPU fault rate of the data. Finally, the normalization operation is performed on all the MAE data, so that the probability of all data generating CPU fault is obtained.

[0067] ​Because the self-encoder model is trained with non-CPU fault data, and the loss function is the mse value of the 138-length vector of the input and output, the self-encoder model learns the data performance characteristics of the non-CPU fault data. Similarly, because the CPU fault data is not learned, the mse value is relatively high. When the non-CPU fault data is inferred through the self-encoder model, the MAE value of the generated vector and the input vector tends to be 0. Similarly, when the CPU fault data is inferred through the self-encoder model, the MAE value of the generated vector and the input vector will be larger.

[0068] S2.4, constructing the device fault self-encoder model, the memory fault self-encoder model and the network fault self-encoder model according to the method of steps S2.1-S2.3, and calculating the probability of each sample in the fault feature sample to generate a device fault, the probability to generate a memory fault and the probability to generate a network fault;

[0069] By sequentially inferring the full amount of data in step 1.3 through multiple fault self-encoder models, the device fault probability, CPU fault probability, memory fault probability and network fault probability of the full amount of data can be obtained. Each model infers the full amount of data, which not only quantifies the potential and high-risk faults, but also mines the factors causing the faults, so as to finally make the fault judgment more accurate and solve the problem that the traditional classification method cannot make a good prediction of the device fault due to the small number of fault devices.

[0070] S3, serializing all the fault probabilities obtained in step S2, and performing feature engineering processing on the fault feature samples in step S1.3 to obtain a fault feature vector F, and using an LSTM network model, a DeepAR model and a linear classifier to construct a fault time series prediction model, including the following steps:

[0071] S3.1, serializing all the fault probabilities obtained in step S2 to obtain a CPU fault probability sequence, a memory fault probability sequence, a network fault probability sequence and a device fault probability sequence, and concatenating all the fault probability sequences with historical fault probabilities to obtain a full amount of fault probability sequence;

[0072] The full-fault probability sequence includes historical fault probabilities and current fault probabilities, and finally the following data is obtained through data splicing: full-fault probability sequence ARR={x, y}, wherein x=[[cpu fault probability, memory fault probability, device fault probability, network fault probability], [cpu fault probability, memory fault probability, device fault probability, network fault probability],..., [cpu fault probability, memory fault probability, device fault probability, network fault probability]] are historical fault probabilities calculated in step S2, and y=[[cpu fault probability, memory fault probability, device fault probability, network fault probability]] is a group of current predicted fault probabilities calculated in step S2.

[0073] S3.2, the fault feature samples in step S1.3 are subjected to feature engineering processing to obtain a fault feature vector F, and the full-fault probability sequence is updated by using the fault feature vector F and the fault feature samples;

[0074] The feature engineering processing includes feature selection, bucketing, discretization, embedding and the like, and finally a feature vector with a length of 561 beneficial to deep model training is obtained. For example, feature selection is performed by using chi-square, and features such as fault date that affect the accuracy of the model can be removed; by means of the bucketing operation, quantified data such as the number of CPUs and the size of memory can be segmented into different intervals, and finally the training process of the model is accelerated; by one-hot encoding of the classification data [cpu model, province where the device is located, city where the device is located, device operator, service id], the classification that cannot be understood by the model can be converted into codes; by embedding, the embedding operation is performed on the two id type attributes of service id and device id, and the model accuracy can be improved by about 3%. In this embodiment, each operation in the feature engineering processing is prior art, and will not be described in detail.

[0075] The fault feature vector F is correspondingly added to the full-fault probability sequence, and finally an updated full-fault probability sequence ARR'={x', y'} is generated, wherein x'=[[cpu fault probability, memory fault probability, device fault probability, network fault probability, fault feature vector F], [cpu fault probability, memory fault probability, device fault probability, network fault probability, fault feature vector F],..., [cpu fault probability, memory fault probability, device fault probability, network fault probability, fault feature vector F]], and y'=[[cpu fault probability, memory fault probability, device fault probability, network fault probability, whether the device is faulty]].

[0076] S3.3, the updated full-failure probability sequence is respectively intercepted according to the fault type, and the intercepted sequence data is respectively input into an LSTM model to obtain an LSTM_CPU model, an LSTM_memory model and an LSTM_network model, and the updated full-failure probability sequence is input into a DeepAR model to obtain a fault distribution prediction model through optimization training;

[0077] The inputs of the LSTM_CPU model, the LSTM_memory model and the LSTM_network model are obtained by intercepting the distribution positions of each fault type in the updated full-failure probability sequence, for example, the input of the LSTM_CPU model is: input_cpu={x”=[[cpu failure probability], [cpu failure probability],..., [cpu failure probability]], y”=[[cpu failure probability]]}

[0078] In this embodiment, the LSTM_CPU model has the following structure:

[0079] Tier Type Number of nodes Activation function First tier LSTM 32 tanh Second tier LSTM 64 tanh Third tier LSTM 32 tanh Fourth tier Dropout Dropout input ratio 0.19 Fifth tier Fully connected layer 1 relu

[0080] As shown in Figure 4 , it is a training effect diagram of the LSTM_CPU model. The LSTM_memory model, the LSTM_network model and the LSTM_device model have the same structure as the LSTM_CPU model.

[0081] The input of the DeepAR model is the updated full-failure probability sequence, and the output is a vector with a length of 4, i.e., the CPU failure probability, the memory failure probability, the device failure probability and the network failure probability. The mean square error (MSE) is used as a loss function to optimize the DeepAR model. Compared with the four LSTM models, the DeepAR model adds a covariant factor F, so it can better fit the influence of external factors on the failure.

[0082] S3.4, as shown in Figure 2 , all the models obtained in step S3.3 are spliced to obtain a fault time series prediction model, and a linear classifier is used to train the fault vector output by the fault time series prediction model to obtain the weight of each dimension vector in the fault vector;

[0083] Since the output of the DeepAR model is a vector with a length of 4, i.e., the CPU failure probability, the memory failure probability, the device failure probability and the network failure probability, the LSTM_CPU model, the LSTM_memory model and the LSTM_network model output vectors with a length of 1 respectively, therefore, the outputs of the four models are spliced to obtain a vector V_model with a length of 7.

[0084] In this embodiment, the linear classifier is sigmoid. The vector V_model is classified by a sigmoid function classifier with y = [[whether the device is malfunctioning]] to obtain the determination of whether the device is malfunctioning through the sigmoid function classifier, and the weight of the corresponding 7 vector dimensions of the vector V_model is also obtained. The weights and parameters of each subnetwork can be better trained by the splicing plus sigmoid method, and the influence of potential high-risk abnormal data on the final device malfunction can be mined.

[0085] S4, as shown in the figure, the fault vector obtained in step S3 is determined and trained by using the FM model, and finally whether the device will malfunction is obtained. Figure 3

[0086] The determination formula is:

[0087]

[0088] In the formula, y represents the classification target value, y = 0 represents no malfunction, y = 1 represents malfunction, w0 represents the bias, w i represents the weight of the vector at the i-th position in the model input vector, x i represents the vector value at the i-th position in the model input vector, n represents the length of the model input vector, v if represents the vector value at the j-th position in the hidden vector of x i , and k represents the length of the preset hidden vector.

[0089] Compared with the sigmoid method, the FM algorithm adds the content of the second-order cross part, which can better evaluate the influence of the failure probability of each component on the overall failure in the combined case, and can better mine the cascade effect caused by the risk explosion. For example, the sigmoid method can only mine the influence of the cpu failure risk probability on the device failure compared with the memory failure risk probability. The FM algorithm can not only mine the weight size of the cpu failure risk probability, but also mine the influence of the cpu failure risk probability and the memory failure risk probability on the overall failure. In actual scenarios, the device failure probability is often not only the influence of a single component, such as the network stability of the device will have a greater fluctuation when the cpu failure risk is high.

[0090] ​The final classification determination is made through the FM model, and compared with the Sigmoid function output probability or the LR algorithm, the AUC evaluation index can be improved by about 3.7%. In summary, through the application of the above self-encoder, time series prediction, FM and other algorithms in the device fault prediction scene, the device fault can be more accurately predicted. Through this method, business problems such as uncertain factors and cascade effects that cannot be solved by conventional algorithms can also be solved.

[0091] Embodiment 2: A deep learning-based edge cloud fault prediction system, comprising:

[0092] A sample collection module is configured to collect fault information of an edge cloud device, and construct a fault feature sample according to the fault information.

[0093] A fault self-encoding model establishment module is configured to train non-fault data in the fault feature sample based on an unsupervised learning method, and construct a CPU fault self-encoding model, a device fault self-encoding model, a memory fault self-encoding model, and a network fault self-encoding model.

[0094] A fault probability calculation module is configured to calculate the mean absolute error according to the corresponding input vector of the fault feature sample and the output vector of the CPU fault self-encoding model, the device fault self-encoding model, the memory fault self-encoding model, and the network fault self-encoding model, and obtain the probability of CPU fault, the probability of device fault, the probability of memory fault, and the probability of network fault according to the mean absolute error.

[0095] A fault time series prediction model construction module is configured to perform feature engineering processing on the fault feature sample to obtain a fault feature vector, and based on the serialized CPU fault probability, device fault probability, memory fault probability, and network fault probability, use an LSTM network model, a DeepAR model, and a linear classifier to construct a fault time series prediction model for predicting the fault vector.

[0096] A fault prediction module is configured to determine and train the fault vector and weight output by the fault time series prediction model construction module based on an FM model, and obtain the fault information of the edge cloud device.

[0097] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A deep learning-based edge cloud failure prediction method, characterized in that, The method comprises the following steps: S1, collecting fault information of edge cloud devices, and constructing a fault feature sample; S2, using an unsupervised learning method to expand train non-fault data in the fault feature sample obtained in step S1, to obtain a CPU fault auto-encoding model, a device fault auto-encoding model, a memory fault auto-encoding model, and a network fault auto-encoding model, and using mean absolute error to calculate the probability of each sample in the fault feature sample to cause CPU fault, device fault, memory fault, and network fault; S3, serializing the fault probabilities output by all fault auto-encoding models, and performing feature engineering processing on the fault feature sample in step S1 to obtain a fault feature vector F, and using an LSTM network model, a DeepAR model, and a linear classifier to construct a fault time series prediction model for predicting the fault vector; The step S3 comprises the following steps: S3.1, serializing all the fault probabilities obtained in step S2 to obtain a CPU fault probability sequence, a memory fault probability sequence, a network fault probability sequence, and a device fault probability sequence, and splicing all the fault probability sequences with historical fault probabilities to obtain a full-amount fault probability sequence; S3.2, performing feature engineering processing on the fault feature sample in step S1 to obtain a fault feature vector F, and updating the full-amount fault probability sequence using the fault feature vector F and the fault feature sample; S3.3, according to the fault type, the updated full-amount fault probability sequence is cut respectively, and the cut sequence data is input into an LSTM model for training to obtain an LSTM_CPU model, an LSTM_memory model, and an LSTM_network model, and the updated full-amount fault probability sequence is input into a DeepAR model for optimization training to obtain a fault distribution prediction model; S3.4, splicing all the models obtained in step S3.3 to obtain a fault time series prediction model for predicting the fault vector, obtaining a corresponding fault vector based on the updated full-amount fault probability sequence and the fault time series prediction model, and training the fault vector using a linear classifier to obtain the weight of each dimension vector in the fault vector; S4, using an FM model to determine and train the fault vector and the weight obtained in step S3 to obtain whether the device will fail. 2.The deep learning based edge cloud failure prediction method of claim 1, wherein, In step S1, the fault feature sample comprises device ID, service ID, device feature, service feature, fault date, whether the CPU fails, whether the memory fails, whether the network fails, and whether the device fails. 3.The deep learning based edge cloud failure prediction method of claim 1, wherein, The step S2 comprises the following steps: S2.1, filtering the non-CPU fault data from the fault feature sample as positive samples, and using a multivariate Gaussian distribution to filter the positive samples; S2.2, using an auto-encoding model to train and optimize the filtered non-CPU fault data to obtain a CPU fault auto-encoding model for predicting CPU fault probability; S2.3, input the fault feature sample of step S1 into the CPU fault auto-encoding model for testing, calculate the mean absolute error according to the vector length of the input non-CPU fault data and the vector length of the model output, normalize all the mean absolute errors to obtain the probability of the fault feature sample generating CPU fault; S2.4, construct the device fault auto-encoding model, the memory fault auto-encoding model and the network fault auto-encoding model according to the method of steps S2.1-S2.3, and calculate the probability of each sample in the fault feature sample generating device fault, the probability of generating memory fault and the probability of generating network fault. 4.The deep learning based edge cloud failure prediction method of claim 3, wherein, In step S2.1, the filtering of the positive samples using the multivariate Gaussian distribution refers to calculating the density probability of the decile number of each sample as the threshold value , the data with the density probability is filtered out The calculation formula of the density probability is: ； wherein denotes the mean, denotes the number of samples, denotes the sample, denotes the covariance. 5.The deep learning based edge cloud failure prediction method of claim 3, wherein, In step S2.2, the hidden layer of the auto-encoding model is three layers, and the node numbers of the three layers are <64, 32, 64> respectively. 6.The deep learning based edge cloud failure prediction method of claim 1, wherein, In step S3.3, the model structure of the LSTM_CPU model, the LSTM_memory model and the LSTM_network model all includes a first LSTM, a second LSTM, a third LSTM, a Dropout layer and a full connection layer connected in sequence. 7.The deep learning based edge cloud failure prediction method of claim 1, wherein, In step S4, the determination formula is: ； In the formula, represents a fault classification target value, represents a deviation, represents a weight of a vector at a th position in a model input vector, represents a vector value at a th position in a model input vector, represents a length of a model input vector, represents a vector value at a th position in a hidden vector of a preset hidden vector. 8.A deep learning based edge cloud failure prediction system, characterized in that, Comprising: a sample collection module for collecting fault information of edge cloud devices and constructing fault feature samples according to the fault information; a fault auto-encoding model establishment module for training non-fault data in the fault feature sample based on an unsupervised learning method, and constructing a CPU fault auto-encoding model, a device fault auto-encoding model, a memory fault auto-encoding model and a network fault auto-encoding model; a fault probability calculation module for calculating the mean absolute error according to the corresponding input vector of the fault feature sample and the output vector of the CPU fault auto-encoding model, the device fault auto-encoding model, the memory fault auto-encoding model and the network fault auto-encoding model, and obtaining the probability of CPU fault, the probability of generating device fault, the probability of generating memory fault and the probability of generating network fault according to the mean absolute error; a fault time series prediction model construction module for performing feature engineering processing on the fault feature sample to obtain a fault feature vector, and constructing a fault time series prediction model for predicting the fault vector based on the serialized CPU fault probability, device fault probability, memory fault probability and network fault probability using an LSTM network model, a DeepAR model and a linear classifier; a fault prediction module for determining and training the fault vector and the weight output by the fault time series prediction model construction module based on the FM model, and obtaining the fault information of the edge cloud device; The construction method of the fault time series prediction model for predicting the fault vector comprises the following steps: S3.1, sequence all the fault probabilities obtained by the fault probability calculation module to obtain a CPU fault probability sequence, a memory fault probability sequence, a network fault probability sequence and a device fault probability sequence, and splice all the fault probability sequences with historical fault probabilities to obtain a full-amount fault probability sequence; S3.2, perform feature engineering processing on the fault feature sample in the sample collection module to obtain a fault feature vector F, and update the full-amount fault probability sequence using the fault feature vector F and the fault feature sample; S3.3, respectively, according to the fault type, the updated full fault probability sequence is intercepted, the intercepted sequence data is respectively input into the LSTM model for training to obtain the LSTM_CPU model, the LSTM_memory model and the LSTM_network model, and the updated full fault probability sequence is input into the DeepAR model for optimization training to obtain the fault distribution prediction model; S3.4, all the models obtained in step S3.3 are spliced to obtain a fault time series prediction model for predicting a fault vector, a corresponding fault vector is obtained based on the updated full fault probability sequence and the fault time series prediction model, and a linear classifier is used to train the fault vector to obtain the weight of each dimension vector in the fault vector.

Images