Server cluster detection method and device, equipment, storage medium and program product

Abstract

The invention relates to a server cluster detection method and device, equipment, a storage medium and a program product. The method comprises the following steps: training a to-be-trained index generation model by using a historical index feature set of a server cluster to obtain a normal index generation model, and training the index generation model by using the historical index feature set to obtain an abnormal index generation model; generating sample normal index characteristics of the server cluster through a normal index generation model, and generating sample abnormal index characteristics of the server cluster through an abnormal index generation model; training a to-be-trained health degree detection model by using the sample normal index features and the sample abnormal index features to obtain a pre-trained health degree detection model; and training a pre-trained health degree detection model by using the historical normal index features and the historical abnormal index features of the server cluster to obtain a target health degree detection model. By adopting the method, the detection accuracy can be improved.

Description

Technical Field

[0001] This application relates to the fields of machine learning and computer technology, and particularly to a server cluster detection method, device, equipment, storage medium, and program product. Background Art

[0002] With the continuous popularization and development of cloud computing technology, server clusters, such as cloud computing server clusters, have become one of the important infrastructures for many enterprises and institutions. Server cluster health detection can help enterprises discover potential faults or hidden problems in the server cluster, and take timely maintenance and management measures to avoid problems such as downtime or data loss, ensure the continuity and availability of services, improve security, reduce operation and maintenance costs, improve customer satisfaction, and thus achieve the sustainable development of enterprises. Health detection is crucial for ensuring the stability of servers.

[0003] In traditional technologies, monitoring tools or agent programs are used to collect key metrics of the server cluster, and the collected metrics are processed using methods such as thresholds or averages, and the health is detected based on the processing results.

[0004] However, the accuracy of traditional health detection is insufficient, and it is necessary to provide a solution that can improve the detection accuracy. Summary of the Invention

[0005] Based on this, in view of the above technical problems, it is necessary to provide a server cluster detection method, device, computer equipment, computer-readable storage medium, and computer program product that can improve the detection accuracy.

[0006] On the one hand, this application provides a server cluster detection method, including: training a to-be-trained metric generation model using a historical metric feature set of the server cluster to obtain a normal metric generation model, and training the metric generation model using the historical metric feature set to obtain an abnormal metric generation model; generating sample normal metric features of the server cluster through the normal metric generation model, and generating sample abnormal metric features of the server cluster through the abnormal metric generation model; training a to-be-trained health detection model using the sample normal metric features and the sample abnormal metric features to obtain a pre-trained health detection model; training the pre-trained health detection model using the historical normal metric features and historical abnormal metric features of the server cluster to obtain a target health detection model, and the target health detection model is used to detect the health of the server cluster.

[0007] On the other hand, the present application also provides a server cluster detection device, including: a generation model obtaining module, configured to train an index generation model to be trained by using a historical index feature set of the server cluster to obtain a normal index generation model, and train the index generation model by using the historical index feature set to obtain an abnormal index generation model; an index feature generation module, configured to generate sample normal index features of the server cluster through the normal index generation model, and generate sample abnormal index features of the server cluster through the abnormal index generation model; a first training module, configured to train a health detection model to be trained by using the sample normal index features and the sample abnormal index features to obtain a pre-trained health detection model; a second training module, configured to train the pre-trained health detection model by using the historical normal index features and historical abnormal index features of the server cluster to obtain a target health detection model, and the target health detection model is used to detect the health of the server cluster.

[0008] On the other hand, the present application also provides a computer device, including a memory and a processor, where the memory stores a computer program, and when the processor executes the computer program, the steps in the above-mentioned server cluster detection method are implemented.

[0009] On the other hand, the present application also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps in the above-mentioned server cluster detection method are implemented.

[0010] On the other hand, the present application also provides a computer program product, including a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned server cluster detection method are implemented.

[0011] The above-mentioned server cluster detection method, device, computer device, computer-readable storage medium and computer program product use the historical metric feature set of the server cluster to train the metric generation model to be trained, obtain a normal metric generation model, and use the historical metric feature set to train the metric generation model to obtain an abnormal metric generation model. Generate the sample normal metric features of the server cluster through the normal metric generation model, and generate the sample abnormal metric features of the server cluster through the abnormal metric generation model. Use the sample normal metric features and sample abnormal metric features to train the health detection model to be trained, obtain a pre-trained health detection model, and use the historical normal metric features and historical abnormal metric features of the server cluster to train the pre-trained health detection model to obtain a target health detection model. The target health detection model is used to detect the health of the server cluster. On the one hand, through the trained normal metric generation model and abnormal metric generation model, sample normal metric features and sample abnormal metric features can be generated. The sample normal metric features can not only simulate the metric features of the server cluster under normal operation, but also have clear labels (without manual annotation). Similarly, the sample abnormal metric features can not only simulate the metric features of the server cluster under abnormal operation, but also have clear labels (without manual annotation), so that there can be sufficient training samples to train the model, improving the generalization ability of the model and helping to improve the accuracy of the health detection model; on the other hand, first use the generated samples to train the health detection model, and then use the real historical normal metric features and historical abnormal metric features to train the health detection model, which can further improve the accuracy of the health detection model; thus, the accuracy of server cluster health detection can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments of the present application or related technologies. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0013] Figure 1 It is an application environment diagram of the server cluster detection method in some embodiments;

[0014] Figure 2 It is a flowchart of the server cluster detection method in some embodiments;

[0015] Figure 3 It is a network structure diagram of the DAE model in some embodiments;

[0016] Figure 4ASchematic diagram of a training metric generation model in some embodiments;

[0017] Figure 4B Schematic diagram of a training metric generation model in other embodiments;

[0018] Figure 5 Schematic diagram of a training health detection model in some embodiments;

[0019] Figure 6 Schematic diagram of a training health detection model in other embodiments;

[0020] Figure 7 Schematic diagram of a training health detection model in still other embodiments;

[0021] Figure 8 Structural block diagram of a server cluster detection device in some embodiments;

[0022] Figure 9 Internal structure diagram of a computer device in some embodiments;

[0023] Figure 10 Internal structure diagram of a computer device in other embodiments. Detailed implementation manners

[0024] In order to make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0025] The server cluster detection method provided by the embodiments of the present application can be applied to an application environment as shown in Figure 1 wherein, the application environment includes a computer device 102 and a server cluster 104. Among them, the computer device 102 communicates with the server cluster 104 through a network.

[0026] Specifically, the computer device 102 uses the historical metric feature set of the server cluster to train the metric generation model to be trained, obtaining a normal metric generation model, and uses the historical metric feature set to train the metric generation model, obtaining an abnormal metric generation model. The computer device 102 generates sample normal metric features of the server cluster through the normal metric generation model, and generates sample abnormal metric features of the server cluster through the abnormal metric generation model. The computer device 102 uses the sample normal metric features and the sample abnormal metric features to train the health detection model to be trained, obtaining a pre-trained health detection model. The computer device 102 uses the historical normal metric features and historical abnormal metric features of the server cluster to train the pre-trained health detection model, obtaining a target health detection model, and the target health detection model is used to detect the health of the server cluster.

[0027] Among them, the computer device can use the sample normal metric features and the sample abnormal metric features to perform semi-supervised learning on the health detection model to be trained, obtaining a pre-trained health detection model. Among them, semi-supervised learning (SSL) is a learning paradigm in the field of machine learning that lies between supervised learning and unsupervised learning. In semi-supervised learning, we simultaneously use labeled supervised data and unlabeled unsupervised data for model training to improve the performance and generalization ability of the learning algorithm. In traditional supervised learning, we need a large number of labeled training samples to train the model. However, the acquisition of labeled data often requires manual operation or expert knowledge, with a high cost. In contrast, unsupervised learning can use unlabeled data for model training, but it has less control and guidance over the learning mode. Semi-supervised learning makes full use of the supervision information of labeled data and the additional information of unlabeled data. It uses the class information of labeled data to guide the clustering or classification of unlabeled data, thus providing richer training signals. By combining the learning of labeled and unlabeled data, semi-supervised learning can reduce the need for labeled data, lower the cost of manual labeling, and can better utilize the implicit structure and distribution characteristics in the data to improve the generalization ability of the model. Semi-supervised learning has a wide range of application scenarios in practical applications, such as image classification, text classification, anomaly detection and other fields. Semi-supervised learning provides a powerful tool and method for machine learning tasks by effectively using limited labeled data and a large amount of unlabeled data.

[0028] Among them, the health of the server cluster refers to the measurement of the overall health and performance status of the server cluster under normal operating conditions. It is used to evaluate the stability, reliability, and working efficiency of the server cluster, and provides a comprehensive index of the current state of the server cluster. Its indicators and monitoring data include but are not limited to the following aspects: hardware health, resource utilization, error rate and failure rate, CPU (Central Processing Unit) usage, memory usage, disk usage, or network traffic, etc.

[0029] Among them, the index generation model to be trained can be a generative adversarial model. A standard generative adversarial model consists of two parts, a generative model and a discriminative model. The task of the generative model is to generate samples that seem natural, real, and similar to the original data. The role of the discriminative model is to determine how real the sample is. The training process also adopts a semi-supervised method and uses the backpropagation mechanism to achieve the dynamic game between the two modules. Finally, the discriminative result of the discriminative model for the sample reaches 50%, maximizing the interests of both parties.

[0030] Among them, the computer device 102 can be a terminal or a server. The computer device 102 can be but is not limited to a desktop computer, a laptop computer, a smart phone, or a tablet computer. Alternatively, the computer device 102 can be an independent physical server, or a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services. The cloud server is used to provide basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. The computer device 102 and the server cluster 104 can be directly or indirectly connected through wired or wireless communication methods, which are not limited in this application. The servers in the server cluster can be cloud servers.

[0031] In some embodiments, as Figure 2 shown, a server cluster detection method is provided. This method can be executed by a terminal or a server, or jointly executed by a terminal and a server. Taking the computer device 102 in Figure 1 as an example, the method includes the following steps 202 to 208. Among them:

[0032] Step 202: Use the historical index feature set of the server cluster to train the index generation model to be trained to obtain a normal index generation model, and use the historical index feature set to train the index generation model to obtain an abnormal index generation model.

[0033] Among them, the metric features are used to reflect the running state of the server cluster, and the running state can be divided into normal and abnormal. The metric features are obtained by processing the metric data collected from the server cluster. The historical metric features are obtained by processing the metric data collected from the server cluster at historical moments. The data processing includes at least one of, but is not limited to, preprocessing or normalization processing.

[0034] The metric data can include metric values under multiple metrics. The metrics can be, but are not limited to, the average number of threads in the running queue within a preset time interval, CPU utilization, remaining disk space, disk I / O (Input / Output) rate, or memory usage, etc. The average number of threads in the running queue within a preset time interval is used to reflect the busyness of the CPU. The larger the average number of threads, the longer the waiting time for time slice allocation and the lower the task processing efficiency. CPU utilization is used to reflect the usage and consumption of the CPU. The remaining disk space is used to reflect the time during which the server cluster can perform log writing and maintain system I / O during operation. The disk I / O rate is used to reflect the interaction rate between the server cluster and external systems such as databases. The memory usage is used to reflect the running ability of the server cluster.

[0035] The historical metric feature set includes multiple historical metric features. Different historical metric features are generated based on the metric data collected from the server cluster at different historical moments. Some of the historical metric features in the historical metric feature set are marked with health labels, and the health labels are normal labels or abnormal labels. If a historical metric feature is generated based on the metric data collected when the server cluster is running normally, the health label of the historical metric feature is a normal label. If a historical metric feature is generated based on the metric data collected when the server cluster is running abnormally, the health label of the historical metric feature is an abnormal label.

[0036] The metric generation model to be trained can be a neural network model, and the metric generation model to be trained includes an encoder, a decoder, and a discriminator. The normal metric generation model may not include a discriminator. The metric generation model can be, but is not limited to, a VAE-GAN model, a VAE-WGAN model, etc. Among them, VAE (Variational Autoencoder) is a variational autoencoder, and GAN (Generative Adversarial Network) is a generative adversarial network. The VAE-GAN model is a hybrid generative model that combines a variational autoencoder and a generative adversarial network. WGAN is the abbreviation of Wasserstein GAN, which evaluates the Wasserstein distance between the generated samples and the real samples through the discriminator to optimize the generator. The Wasserstein distance is also known as the Earth-Mover (EM) distance, which is a measure of the difference between two probability distributions.

[0037] Step 204: Generate the sample normal metric features of the server cluster through the normal metric generation model, and generate the sample abnormal metric features of the server cluster through the abnormal metric generation model.

[0038] Among them, the sample normal metric features and the sample abnormal metric features are pseudo data. The sample normal metric features can not only simulate the metric features of the server cluster under normal operating conditions, but also automatically generate normal labels for the sample normal metric features (without manual annotation). Similarly, the sample abnormal metric features can not only simulate the metric features of the server cluster under abnormal operating conditions, but also automatically generate abnormal labels for the sample abnormal metric features (without manual annotation).

[0039] Specifically, the computer device can generate random noise, input the random noise into the normal metric generation model, and obtain the sample normal metric features generated by the normal metric generation model. The computer device can generate random noise, input the random noise into the abnormal metric generation model, and obtain the sample abnormal metric features generated by the abnormal metric generation model.

[0040] Step 206: Use the sample normal metric features and the sample abnormal metric features to train the health detection model to be trained, and obtain the pre-trained health detection model.

[0041] Among them, the health detection model to be trained includes an encoding network, a decoding network, and a classification network. The health detection model can be a model based on AE (Auto-encoder) or DAE (Denoising Autoencoder). The AE model is an unsupervised model, and its overall structure includes two parts: an encoding network and a decoding network. The encoding network and the decoding network are trained using unlabeled data. By minimizing the error through backpropagation, the model output is made as close as possible to the model input. An ideal autoencoding network is equivalent to an identity function. Since the intermediate layer encoding obtained from the data through the encoding network can be decoded by the decoding network without loss to produce the output data, it can be considered that the intermediate layer is a representation with a lower dimension that contains all the features of the input data. Therefore, an autoencoding network is usually used as a feature extraction model. The encoding process of AE can be expressed as: , where represents the weights of the encoding network, represents the bias parameters of the encoding network, is the activation function of the encoding network, is the input data of the encoding network (also the input data of AE), is the output data of the encoding network. The decoding process of AE can be expressed as: , where represents the weights of the decoding network, represents the bias parameters of the decoding network, is the activation function of the decoding network, is the input data of the decoding network, is the output data of the decoding network. The purpose of AE is to minimize the difference value between the input data and the output . The difference value can be described using the mean square error. For example, the difference value (mean square error) can be expressed as:

[0042]

[0043] where represents the number of samples, is the mean square error value for each training, and the AE model can be trained by minimizing . Among them, represents the i-th sample in this training, represents the input in this training and the output result of the AE model.

[0044] The DAE model continues the overall concept of the AE model. Different from the AE model, the DAE actively adds random noise to the sample data during the training process, and trains the model in this way to improve the robustness of the information in the intermediate encoding layer to the input data. For the input data containing noise, the AE model is affected by the noise and the intermediate layer encoding cannot fully reflect the input information, while the DAE model can reduce the dependence on the input information and tolerate the noise contained in the input information. As Figure 3 shown, a network structure diagram of a DAE model is provided, where, ( ) are the features obtained by processing the data ( ), ( ) are the features of the previous intermediate layer, and ( ) is the output result (reconstruction result).

[0045] Specifically, the computer device can use the sample normal index features and sample abnormal index features to first perform unsupervised training on the health detection model to be trained, and then perform supervised training to obtain a pre-trained health detection model.

[0046] Step 208, use the historical normal index features and historical abnormal index features of the server cluster to train the pre-trained health detection model to obtain a target health detection model, where the target health detection model is used to detect the health of the server cluster.

[0047] Among them, in a large computer room, a large number of server clusters are often deployed. Each cluster contains multiple physical nodes, and each node is running services continuously and receiving requests from clients. If the cluster fails, it will seriously affect business continuity. Therefore, cluster health monitoring is a very effective operation and maintenance means, which can help operation and maintenance personnel quickly discover and locate unhealthy clusters and dispose of them in time, and can minimize business losses to the greatest extent.

[0048] The historical normal index features and historical abnormal index features are real features, not generated by the model. The historical normal index features and historical abnormal index features are both marked with health labels. The health label of the historical normal index features is a normal label, and the health label of the historical abnormal index features is an abnormal label. The historical normal index features and historical abnormal index features can be the features in the historical index feature set or not.

[0049] In some embodiments, after the computer device trains the target health detection model, it can deploy the target health detection model to the server cluster, or other devices can deploy the target health detection model trained by the computer device to the server cluster.

[0050] In some embodiments, a computer device or a server cluster collects the metric features of the server cluster in real time to obtain real-time metric features, inputs the real-time metric features into a target health detection model, and uses the encoding network encoding and classification network classification in the target health detection model to obtain the real-time health degree output by the target health detection model. This real-time health degree is used to reflect the current health degree of the server cluster.

[0051] In some embodiments, a computer device can deploy an automatic execution script in a cluster in a computer room, regularly output cluster health information, and input the information into a target health detection model in real time, and judge the health degree of the cluster according to the output result, and send an alarm message in time when it is found that the cluster is abnormal.

[0052] Generally, the supervised learning method for evaluating the health degree of a server cluster has strong dependence on data, requires a large amount of server cluster metric data including known normal and abnormal data, and the generalization ability of the supervised model is limited and the sharing degree is low. A specifically trained model can usually only detect a specific server cluster, and even cannot respond well to abnormal metrics outside the labeled data.

[0053] In the above server cluster detection method, the historical metric feature set of the server cluster is used to train the metric generation model to be trained, obtaining a normal metric generation model, and the historical metric feature set is used to train the metric generation model to obtain an abnormal metric generation model. The sample normal metric features of the server cluster are generated through the normal metric generation model, and the sample abnormal metric features of the server cluster are generated through the abnormal metric generation model. The sample normal metric features and the sample abnormal metric features are used to train the health detection model to be trained, obtaining a pre-trained health detection model. The historical normal metric features and historical abnormal metric features of the server cluster are used to train the pre-trained health detection model, obtaining a target health detection model. The target health detection model is used to detect the health of the server cluster. On the one hand, through the trained normal metric generation model and abnormal metric generation model, sample normal metric features and sample abnormal metric features can be generated. The sample normal metric features can not only simulate the metric features of the server cluster under normal operation conditions, but also have clear labels (without manual annotation). Similarly, the sample abnormal metric features can not only simulate the metric features of the server cluster under abnormal operation conditions, but also have clear labels (without manual annotation), so that there can be sufficient training samples to train the model, improving the generalization ability of the model and helping to improve the accuracy of the health detection model. On the other hand, first using the generated samples to train the health detection model and then using the real historical normal metric features and historical abnormal metric features to train the health detection model can further improve the accuracy of the health detection model. Thus, the accuracy of server cluster health detection can be improved.

[0054] In some embodiments, using the historical metric feature set of the server cluster to train the metric generation model to be trained, obtaining a normal metric generation model includes: adding random noise to the first historical metric features to obtain first updated historical metric features, inputting the first updated historical metric features into the metric generation model to be trained, and obtaining first reconstructed features through encoder encoding and decoder decoding; based on the first difference value between the first reconstructed features and the first historical metric features, training the metric generation model to be trained to obtain a preliminarily trained metric generation model; inputting the second historical metric features into the discriminator to obtain a first discrimination result; adding random noise to the second historical metric features to obtain second updated historical metric features, inputting the second updated historical metric features into the preliminarily trained metric generation model, obtaining second reconstructed features through encoder encoding and decoder decoding, and having the discriminator discriminate the second reconstructed features to obtain a second discrimination result, and the health label of the second historical metric features is a normal label; based on the second difference value between the second reconstructed features and the second historical metric features, the first discrimination result, and the second discrimination result, training the preliminarily trained metric generation model to obtain a normal metric generation model.

[0055] Among them, the first historical index feature is any historical index feature in the set of historical index features. The second historical index feature is the historical index feature in the set of historical index features that is marked with a health label and the health label is a normal label. It should be noted that in this application, random noise needs to be added at multiple places, and the random noise added each time is randomly generated, and is not limited to the same random noise. The authenticity label is used to distinguish between pseudo-features (i.e., features generated by the model) and real features (i.e., index features generated according to index data). Since the second historical index feature is a real feature, the authenticity label of the second historical index feature is "true" (which can be represented by the number 1 for "true"). Obviously, the authenticity label of the second reconstructed feature is "false" (which can be represented by the number 0 for "false").

[0056] In some embodiments, the structural schematic diagram of the index generation model to be trained includes an encoder, a decoder, and a discriminator. Among them, the encoder can be used to map the input index features to the mean and variance of the latent space. The encoder can be a fully connected network. For example, it can be composed of three fully connected layers. The first fully connected layer is responsible for performing downsampling operations on the input multi-dimensional data, mapping the input data to a low-dimensional space to extract deeper features; the following two fully connected layers then map the downsampled sample data of the cluster health to an even lower dimension to represent the mean ( ), and variance ( ). The encoder corresponds to a loss function , and this loss function uses the KL (Kullback-Leibler Divergence) divergence. For example,

[0057]

[0058] Among them, is used to measure the difference between the latent space distribution (implicit vector distribution) generated by the encoder and the standard normal distribution, so that the implicit vector distribution is closer to the standard normal distribution. Its expression is as follows. Among them, is the mean output by the encoder, and is the variance output by the encoder.

[0059] The decoder can be understood as a generation network, mainly used to sample latent vectors from the latent space of the encoder to generate new data. The decoder can perform two fully connected operations on the implicit variables output by the encoder, but the parameter settings are opposite to those of the encoder.

[0060] The loss function of the decoder includes two parts: reconstruction loss and adversarial loss. Among them, the adversarial loss uses the Wasserstein distance. The loss function of the decoder , where is the reconstruction loss, is the original data (e.g., the first historical metric feature), is the reconstructed data (e.g., the first reconstructed feature). is the adversarial loss. represents the output result of the decoder, i.e., , is the latent vector generated by the encoder. represents the result of the discriminator's judgment on authenticity.

[0061] The discriminator can have convolutional layers and fully connected layers. For example, the discriminator can consist of two convolutional layers and three fully connected layers. The last output layer is a fully connected layer without an activation function. It should be noted that because WGAN is used, this discriminator network needs to satisfy the Lipschitz continuity condition, which is usually achieved through weight clipping, restricting the weight range of each layer to be between [-0.01, 0.01].

[0062] The discriminator is used to measure the Wasserstein distance between the original data (e.g., the first historical metric feature) and the generated data (i.e., the reconstructed data such as the first reconstructed feature), and to judge the authenticity of the generated data through the Wasserstein distance. The loss function of the discriminator is, for example:

[0063]

[0064] where is the output result of the discriminator after inputting the original data into the discriminator.

[0065] During the process of training the metric generation model, the difference degree between the generated data and the real data can be calculated, and this can be used as an evaluation metric to evaluate the model performance. According to the evaluation results, the model can be optimized by adjusting the network layers or hyperparameters, etc., so that the finally generated server cluster data is closer to the real server cluster data.

[0066] Specifically, from the metric generation model to be trained to the preliminarily trained metric generation model, it is unsupervised training. From the preliminarily trained metric generation model to the normal metric generation model, since the discriminator is involved, it reflects the process of true-false game, so it can be considered a kind of supervised training. Thus, overall, it can be considered a kind of semi-supervised training.

[0067] Such as Figure 4AAs shown, during the unsupervised training process, the encoder encodes the first updated historical metric feature to obtain a latent vector, and the decoder decodes the latent vector to obtain a first reconstructed feature. The reconstruction loss value can be generated according to the above reconstruction loss calculation formula and the first difference value. The parameters of the encoder and decoder in the metric generation model to be trained are adjusted according to the reconstruction loss value. Of course, the encoding loss value can also be generated according to the loss function corresponding to the encoder, the parameters of the encoder and decoder in the metric generation model to be trained are adjusted according to the reconstruction loss value, and the parameters of the encoder are updated according to the encoding loss value to obtain a preliminarily trained metric generation model. Among them, there can be multiple first historical metric features, and a preliminarily trained metric generation model can be obtained through multiple iterative trainings.

[0068] As Figure 4B shown, during the supervised training process, the encoder encodes the second updated historical metric feature to obtain a latent vector, the decoder decodes the latent vector to obtain a second reconstructed feature, and the second reconstructed feature is input into the discriminator to obtain a discrimination result. The computer device can substitute the second difference value and the second discrimination result into the loss function of the decoder to obtain the loss value of the decoder, and substitute the second discrimination result and the first discrimination result into the loss function of the discriminator to obtain the loss value of the discriminator. Then, the parameters of the decoder and encoder are updated according to the loss value of the decoder, and the parameters of the discriminator are updated according to the loss value of the discriminator. After iterative training using multiple first historical metric features and multiple second historical metric features, a normal metric generation model is obtained. The normal metric generation model can only include an encoder and a decoder, and of course, it can also include a discriminator. The training process of the abnormal metric generation model is similar and will not be elaborated here.

[0069] In this embodiment, by performing semi-supervised training on the metric generation model, the similarity between the features generated by the normal metric generation model and the real metric features is improved.

[0070] In some embodiments, generating sample normal metric features of the server cluster through the normal metric generation model includes: generating random noise, inputting the random noise into the normal metric generation model, and obtaining the output result of the decoder in the normal metric generation model; using the output result of the decoder in the normal metric generation model as the sample normal metric features of the server cluster.

[0071] Among them, the random noise can randomly generate noise, which is not necessarily the same noise as the random noise in the other embodiments of this application.

[0072] In this embodiment, by inputting random noise into the normal index generation model, sample normal index features can be obtained, so that the health label of the sample normal index features can be determined as the normal label, which facilitates the rapid generation of labeled samples and provides convenience for training the health detection model.

[0073] In some embodiments, the health detection model to be trained is trained using sample normal index features and sample abnormal index features to obtain a pre-trained health detection model, including: adding random noise to the first sample normal index features to obtain first updated sample normal index features; adding random noise to the first sample abnormal index features to obtain first updated sample abnormal index features; inputting the first updated sample normal index features into the health detection model to be trained, and obtaining a first output result through encoding by the encoding network and decoding by the decoding network; inputting the first updated sample abnormal index features into the health detection model to be trained, and obtaining a second output result through encoding by the encoding network and decoding by the decoding network; adjusting the parameters of the health detection model to be trained based on the third difference value between the first output result and the first sample normal index features and the fourth difference value between the second output result and the first sample abnormal index features to obtain a preliminarily trained health detection model; and obtaining a pre-trained health detection model based on the preliminarily trained health detection model.

[0074] Among them, when training the health detection model, multiple sample normal index features and multiple sample abnormal index features can be used, where multiple means at least two. The first sample normal index feature is any one of the multiple sample normal index features. The first sample abnormal index feature is any one of the multiple sample abnormal index features. The health detection model to be trained includes an encoding network and a decoding network, and also includes a decoding network.

[0075] Specifically, as Figure 5 shown, a schematic diagram for training the health detection model to be trained is provided. The computer device encodes the first updated sample normal index features using the encoding network to obtain a first encoded feature, and decodes the first encoded feature using the decoding network to obtain a first output result. Similarly, a second output result can be obtained.

[0076] In some embodiments, there can be multiple first-sample normal-index features and multiple second-sample normal-index features, so that multiple third difference values and multiple fourth difference values can be obtained. The computer device can use the multiple third difference values and multiple fourth difference values to generate a first model loss value. For example, the multiple third difference values and multiple fourth difference values can be averaged, and the calculated average value can be used as the first model loss value. That is, the mean squared error can be used to calculate the first model loss value. The first-sample normal-index features and the second-sample normal-index features are respectively a sample during training. Then, the computer device adjusts the parameters of the encoding network and the decoding network in the health detection model to be trained by using the first model loss value, and obtains a preliminarily trained health detection model.

[0077] In this embodiment, based on the third difference value between the first output result and the first-sample normal-index features, and the fourth difference value between the second output result and the first-sample abnormal-index features, the parameters of the health detection model to be trained are adjusted to obtain a preliminarily trained health detection model, realizing unsupervised training (i.e., without using labels) of the health detection model. Through unsupervised training, the encoding features generated by the encoding network in the model are more accurate, which helps to improve the accuracy of model detection. Usually, the definition of anomaly detection in server health detection work is an anomaly in a statistical sense. A change in the detection index does not mean that the server cluster loses its service ability, and the reliability of health detection is poor. The training data during the training process usually also contains unknown abnormal data, resulting in the model lacking sufficient robustness. In this embodiment, by training the model with samples containing noise, the robustness of the model can be improved. Since the generated data (the first-sample normal-index features and the first-sample abnormal-index features) is not real cluster health data, the generated data will contain more noise. Using this data for pre-training can further increase the robustness of the model and improve the model's fault tolerance ability.

[0078] In some embodiments, obtaining a pre-trained health detection model based on the preliminarily trained health detection model includes: adding random noise to the second-sample normal-index features to obtain second-updated sample normal-index features; adding random noise to the second-sample abnormal-index features to obtain second-updated sample abnormal-index features; inputting the second-updated sample normal-index features into the preliminarily trained health detection model, encoding through the encoding network and classifying through the classification network to obtain a first classification result; inputting the second-updated sample abnormal-index features into the preliminarily trained health detection model, encoding through the encoding network and classifying through the classification network to obtain a second classification result; adjusting the parameters of the preliminarily trained health detection model based on the first classification result and the normal label, and the second classification result and the abnormal index, to obtain a pre-trained health detection model.

[0079] Specifically, as Figure 6As shown, a schematic diagram for training a preliminarily trained health detection model is provided. The computer device encodes the normal index features of the second updated sample using an encoding network to obtain second encoded features, and classifies the second encoded features using a classification network to obtain a first classification result. Similarly, a second classification result can be obtained. Among them, the encoding network can extract a low-dimensional feature representation of the input information, and this feature can be used as a complete expression of the input information. The classification network is a neural network. The classification network is located after the encoding network. The classification network can be composed of neural network layers including a softmax function. After the output result of the encoding network passes through this multi-layer network and the softmax layer, the health detection result of the cluster is obtained.

[0080] Among them, this multi-layer network can be represented as D, and the output of the softmax layer can be represented as , , . is the output data of the encoding network, M and N represent the weight and bias parameters of the classification network, is the activation function. The activation function can be but is not limited to the relu function; the feature D passes through a softmax function to obtain the final detection result. The softmax function is as follows,

[0081]

[0082] Among them, represents the output of the i-th node, C represents the number of output nodes, that is, the detection result categories, such as healthy, abnormal, and unknown, etc. Through the softmax function, the classification output value can be converted into a probability distribution within the range of [0,1] and the sum is 1.

[0083] In some embodiments, since the normal index features of the second sample are generated by a normal index generation model, the health label of the normal index features of the second sample is a normal label. Similarly, the health label of the abnormal index features of the second sample is an abnormal label. The first classification result and the second classification result can be probability values respectively, and the normal label and the abnormal label are also probability values respectively. The computer device can generate a cross-entropy loss value (denoted as the second model loss value) based on the first classification result and the normal label, the second classification result and the abnormal index, in combination with the cross-entropy loss function (the purpose of this loss function is to minimize the error between the classification result output by the classification network and the label). Then, the parameters of the decoding network and the classification network in the preliminarily trained health detection model are adjusted using the second model loss value to obtain a pre-trained health detection model.

[0084] In this embodiment, based on the first classification result and the normal label, and the second classification result and the abnormal index, the parameters of the preliminarily trained health detection model are adjusted to obtain a pre-trained health detection model, so that the health detection model can learn the ability to correctly classify according to the input features.

[0085] In some embodiments, the third historical index feature in the historical index feature set corresponds to a normal label, and the normal label is used to represent that the server cluster is normal. The fourth historical index feature in the historical index feature set corresponds to an abnormal label, and the abnormal label is used to represent that the server cluster is abnormal. The pre-trained health detection model is trained using the historical normal index features and historical abnormal index features of the server cluster to obtain a target health detection model, including: adding random noise to the third historical index feature to obtain a first updated historical index feature; adding random noise to the fourth historical index feature to obtain a second updated historical index feature; inputting the first updated historical index feature into the pre-trained health detection model to obtain a third output result of the decoding network and obtaining a third classification result of the classification network; inputting the second updated historical index feature into the pre-trained health detection model to obtain a fourth output result of the decoding network and obtaining a fourth classification result of the classification network; based on the fifth difference value between the third output result and the third historical index feature, the sixth difference value between the fourth output result and the fourth historical index feature, the third classification result, the normal label, the fourth classification result, and the abnormal label, the parameters of the pre-trained health detection model are adjusted to obtain a target health detection model.

[0086] Among them, as Figure 7 shown, a schematic diagram for training the pre-trained health detection model is provided. According to the third historical index feature, a third output result and a third classification result are obtained. Similarly, a fourth output result and a fourth classification result can be obtained.

[0087] Specifically, the computer device can generate a decoding loss value according to the fifth difference value and the sixth difference value, generate a classification loss value according to the third classification result and the normal label, and the fourth classification result and the abnormal label, and perform weighted summation on the decoding loss value and the classification loss value to obtain a total loss value. For example, , where, is the total loss value, is the decoding loss value (such as the mean square error loss of the denoising autoencoder branch), is the classification loss value (such as the classification cross-entropy loss). represents the loss weight, by adjusting Experiment with weights to obtain appropriate classification guidance weights, making the cluster health detection results more accurate. The computer device can adjust the parameters of the pre-trained health detection model using the total loss value to obtain the target health detection model. The target health detection model can only include an encoding network and a classification network, or it can also include a decoding network.

[0088] In some embodiments, the computer device can adjust the parameters of the encoding network and the decoding network in the pre-trained health detection model using the decoding loss value, and adjust the parameters of the encoding network and the classification network in the pre-trained health detection model using the classification loss value to obtain the target health detection model.

[0089] In this embodiment, based on the fifth difference value between the third output result and the third historical indicator feature, the sixth difference value between the fourth output result and the fourth historical indicator feature, the third classification result and the normal label, and the fourth classification result and the abnormal label, the parameters of the pre-trained health detection model are adjusted to obtain the target health detection model, thereby further improving the detection accuracy (i.e., classification accuracy) of the health detection model.

[0090] In some embodiments, the server cluster detection method provided by this application may include:

[0091] 1. Collect diverse historical data of health monitoring indicators from the monitoring center of the server cluster and label some of the indicator data.

[0092] Specifically, collect historical indicator data from the server cluster monitoring center, including the average number of threads in the queue, CPU utilization, disk space, disk I / O rate, and memory usage within a specific time period. Partially label the collected sample data according to historical failure scenarios.

[0093] 2. Construct a VAE-WGAN data generation model: Use the labeled data and unlabeled data obtained in step 1 to perform semi-supervised training on the data generation model, and separately train a normal operation indicator generation model (i.e., a normal indicator generation model) and an abnormal operation indicator generation model (i.e., an abnormal indicator generation model). The trained models are used to generate sample indicator features (such as sample normal indicator features and sample abnormal indicator features).

[0094] Specifically, (1) Construct the encoder of the VAE-WGAN model. For the given metric data, the encoder can encode it into a deeper latent variable, enabling the model to learn deeper data features while meeting the data dimensionality reduction requirements. (2) Construct the decoding part of the VAE-WGAN model, namely the generation network. The role of this module is to sample latent vectors from the latent space of the encoder to generate new data, generating cluster data similar to the original server cluster. (3) Construct the discriminant network of the VAE-WGAN model. It is used to determine the authenticity of the data generated by the above generation module. (4) Sample dataset augmentation: Use the trained VAE-WGAN model to generate labeled sample data for the subsequent training of the autoencoder.

[0095] 3. Construct a semi-supervised denoising autoencoder (DAE) model as the health detection model for the server cluster, and pre-train the model with the data generated in step 2.

[0096] Specifically, different from the traditional autoencoder model, the DAE model actively adds noise pollution to the input data during the training process, and uses the samples containing noise to reconstruct the original samples, so as to achieve the purpose of learning the essential features of the data.

[0097] 4. Construct a cluster health detection model, and use the real server cluster data obtained in step 1 to continue training the DAE model in step 3, and fine-tune the model parameters through training to obtain a perfect cluster health detection model.

[0098] Specifically, use the data generated by the above generation model and the real sample data collected and labeled in step 1 to train the DAE model. The training process includes pre-training with generated data and fine-tuning training with real data, and finally obtain a server cluster health detection model.

[0099] 5. Health detection of the server cluster: Sample data from the server cluster monitoring center at fixed time intervals, for example, set a 1-minute interval for sampling, input the sampled data into the above-trained DAE model, output the server cluster health result, and give an alarm in time if there is an abnormality.

[0100] Specifically, collect the above metric data from the monitoring center through a scheduled task and input it into the cluster health detection model, evaluate the data result, and give an early warning and troubleshooting in time if there is an abnormal result.

[0101] This application applies the semi-supervised learning concept and method to the server cluster health detection work. The semi-supervised learning method can better utilize unsupervised data, reduce the dependence on data, and to a certain extent, reduce the human and material costs required for labeled data. During the operation of the server cluster, there are many measurable indicators, and different server clusters have inconsistent feedback on the indicators. The autoencoder model can achieve data dimensionality reduction and learn the underlying essential features of the data, which is very suitable for the server cluster health detection field, enhancing the generalization of the health detection model and can be migrated to different server clusters after fine-tuning. In addition, this application also introduces a generative adversarial module and a noise reduction module. The former aims to generate sufficient server cluster health detection index data as unsupervised training samples, and the latter can reduce the robustness impact brought by abnormal indicators of labeled data and unlabeled data during the training process. First, perform data augmentation operations on the server cluster through a generative model such as the VAE-WGAN model to increase data diversity; then, perform semi-supervised training on the server cluster through a health detection model such as the DAE model with added random noise and classification layer; finally, first pre-train the semi-supervised health detection model such as the denoising autoencoder with the data generated by the VAE-WGAN, and then fine-tune the model with real server cluster label data.

[0102] In this application, a VAE-WGAN model can be constructed to generate more virtual data of the server cluster. This model not only improves the generalization ability of the model to a certain extent, but also can alleviate the problem of sample imbalance because in the samples of server cluster health detection, the normal data samples are much higher than the abnormal data. In addition, by using the Wasserstein distance as a metric, the model has higher stability compared to the traditional generative adversarial network. In this application, starting from the semi-supervised learning method, through the denoising autoencoder, not only can the data characteristics of the server cluster be effectively learned, but also multi-dimensional data such as recipe usage rate and disk usage rate obtained from the server cluster can be compressed and dimensionally reduced to extract key features. In this application, the output direction of the encoding module of the denoising autoencoder is divided into two parts. One part is to introduce a classification layer to judge whether the server data is healthy, and the other part is to perform decoding to reconstruct the information of the input data. The addition of random noise reduces the dependence on data during the training process and improves the robustness of the model. In this application, first pre-train the semi-supervised autoencoder model with the pseudo-data generated by a generative model such as VAE-WGAN, and then fine-tune the pre-trained model with real data, which not only reduces the risk of overfitting, but also increases the accuracy of the model.

[0103] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are displayed in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same moment, but can be executed at different moments. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0104] Based on the same inventive concept, an embodiment of the present application also provides a server cluster detection device for implementing the above-mentioned server cluster detection method. The solution provided by this device to solve the problem is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the server cluster detection device provided below can refer to the limitations on the server cluster detection method in the above text, and will not be repeated here.

[0105] In some embodiments, as Figure 8 shown, a server cluster detection device is provided, including: a generation model obtaining module 802, an index feature generation module 804, a first training module 806, and a second training module 808, where:

[0106] The generation model obtaining module 802 is used to train the to-be-trained index generation model with the historical index feature set of the server cluster to obtain a normal index generation model, and train the index generation model with the historical index feature set to obtain an abnormal index generation model.

[0107] The index feature generation module 804 is used to generate sample normal index features of the server cluster through the normal index generation model, and generate sample abnormal index features of the server cluster through the abnormal index generation model.

[0108] The first training module 806 is used to train the to-be-trained health detection model with the sample normal index features and the sample abnormal index features to obtain a pre-trained health detection model.

[0109] The second training module 808 is used to train the pre-trained health detection model with the historical normal index features and the historical abnormal index features of the server cluster to obtain a target health detection model, and the target health detection model is used to detect the health of the server cluster.

[0110] In some embodiments, the generation model obtaining module 802 is further configured to add random noise to the first historical metric feature to obtain a first updated historical metric feature, input the first updated historical metric feature into the metric generation model to be trained, and obtain a first reconstructed feature through encoding by the encoder and decoding by the decoder; based on the first difference value between the first reconstructed feature and the first historical metric feature, train the metric generation model to be trained to obtain a preliminarily trained metric generation model; input the second historical metric feature into the discriminator to obtain a first discrimination result; add random noise to the second historical metric feature to obtain a second updated historical metric feature, input the second updated historical metric feature into the preliminarily trained metric generation model, obtain a second reconstructed feature through encoding by the encoder and decoding by the decoder, and obtain a second discrimination result through discrimination of the second reconstructed feature by the discriminator, where the health label of the second historical metric feature is a normal label; based on the second difference value between the second reconstructed feature and the second historical metric feature, the first discrimination result, and the second discrimination result, train the preliminarily trained metric generation model to obtain a normal metric generation model.

[0111] In some embodiments, the metric feature generation module 804 is further configured to generate random noise, input the random noise into the normal metric generation model, and obtain the output result of the decoder in the normal metric generation model; use the output result of the decoder in the normal metric generation model as the sample normal metric feature of the server cluster.

[0112] In some embodiments, the first training module 806 is further configured to add random noise to the first sample normal metric feature to obtain a first updated sample normal metric feature; add random noise to the first sample abnormal metric feature to obtain a first updated sample abnormal metric feature; input the first updated sample normal metric feature into the health detection model to be trained, and obtain a first output result through encoding by the encoding network and decoding by the decoding network; input the first updated sample abnormal metric feature into the health detection model to be trained, and obtain a second output result through encoding by the encoding network and decoding by the decoding network; based on the third difference value between the first output result and the first sample normal metric feature and the fourth difference value between the second output result and the first sample abnormal metric feature, adjust the parameters of the health detection model to be trained to obtain a preliminarily trained health detection model; based on the preliminarily trained health detection model, obtain a pre-trained health detection model.

[0113] In some embodiments, the first training module 806 is further configured to add random noise to the second sample normal index features to obtain second updated sample normal index features; add random noise to the second sample abnormal index features to obtain second updated sample abnormal index features; input the second updated sample normal index features into the preliminarily trained health detection model, and obtain a first classification result through encoding by the encoding network and classification by the classification network; input the second updated sample abnormal index features into the preliminarily trained health detection model, and obtain a second classification result through encoding by the encoding network and classification by the classification network; based on the first classification result and the normal label, and the second classification result and the abnormal index, adjust the parameters of the preliminarily trained health detection model to obtain a pre-trained health detection model.

[0114] In some embodiments, a normal label corresponds to the third historical index feature in the historical index feature set, and the normal label is used to represent that the server cluster is normal; an abnormal label corresponds to the fourth historical index feature in the historical index feature set, and the abnormal label is used to represent that the server cluster is abnormal; the second training module 808 is further configured to add random noise to the third historical index feature to obtain a first updated historical index feature; add random noise to the fourth historical index feature to obtain a second updated historical index feature; input the first updated historical index feature into the pre-trained health detection model to obtain a third output result of the decoding network and obtain a third classification result of the classification network; input the second updated historical index feature into the pre-trained health detection model to obtain a fourth output result of the decoding network and obtain a fourth classification result of the classification network; based on the fifth difference value between the third output result and the third historical index feature, the sixth difference value between the fourth output result and the fourth historical index feature, the third classification result, the normal label, the fourth classification result, and the abnormal label, adjust the parameters of the pre-trained health detection model to obtain a target health detection model.

[0115] Each module in the above server cluster detection device can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or be independent of it, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above respective modules.

[0116] In some embodiments, a computer device is provided. The computer device can be a server, and its internal structural diagram can be as Figure 9As shown in the figure. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the data involved in the server cluster detection method. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements a server cluster detection method.

[0117] In some embodiments, a computer device is provided. The computer device can be a terminal, and its internal structure diagram can be as Figure 10 shown in the figure. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit, and an input device. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface, the display unit, and the input device are connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, a mobile cellular network, near field communication (Near Field Communication, NFC), or other technologies. When the computer program is executed by the processor, it implements a server cluster detection method. The display unit of the computer device is used to form a visually visible picture, which can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the computer device housing, or an external keyboard, touchpad, or mouse, etc.

[0118] Those skilled in the art can understand that Figure 9 and Figure 10The structure shown is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have a different component arrangement.

[0119] In some embodiments, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above-mentioned server cluster detection method are implemented.

[0120] In some embodiments, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned server cluster detection method are implemented.

[0121] In some embodiments, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above-mentioned server cluster detection method are implemented.

[0122] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in this application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in this application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, artificial intelligence (AI) processors, etc., without limitation.

[0123] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this application.

[0124] The above-described embodiments merely represent several implementation manners of the present application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A server cluster detection method, characterized in that, The method includes: Training an index generation model to be trained using the historical index feature set of a server cluster to obtain a normal index generation model, and training the index generation model using the historical index feature set to obtain an abnormal index generation model; Generating sample normal index features of the server cluster through the normal index generation model, and generating sample abnormal index features of the server cluster through the abnormal index generation model; Training a health detection model to be trained using the sample normal index features and the sample abnormal index features to obtain a pre-trained health detection model; Training the pre-trained health detection model using the historical normal index features and historical abnormal index features of the server cluster to obtain a target health detection model, where the target health detection model is used to detect the health of the server cluster.

2. The method according to claim 1, characterized in that The training of the index generation model to be trained using the historical index feature set of the server cluster to obtain a normal index generation model includes: Adding random noise to the first historical index feature to obtain a first updated historical index feature, inputting the first updated historical index feature into the index generation model to be trained, and obtaining a first reconstructed feature through encoder encoding and decoder decoding; Training the index generation model to be trained based on the first difference value between the first reconstructed feature and the first historical index feature to obtain a preliminarily trained index generation model; Inputting the second historical index feature into a discriminator to obtain a first discrimination result; Adding random noise to the second historical index feature to obtain a second updated historical index feature, inputting the second updated historical index feature into the preliminarily trained index generation model, obtaining a second reconstructed feature through encoder encoding and decoder decoding, and having the discriminator discriminate the second reconstructed feature to obtain a second discrimination result, where the health label of the second historical index feature is a normal label; Training the preliminarily trained index generation model based on the second difference value between the second reconstructed feature and the second historical index feature, the first discrimination result, and the second discrimination result to obtain a normal index generation model.

3. The method according to claim 2, wherein The generating of the sample normal index features of the server cluster through the normal index generation model includes: Generating random noise, inputting the random noise into the normal index generation model, and obtaining the output result of the decoder in the normal index generation model; Taking the output result of the decoder in the normal index generation model as the sample normal index features of the server cluster.

4. The method according to any one of claims 1 to 3, characterized in that The training of the health detection model to be trained using the sample normal index features and the sample abnormal index features to obtain a pre-trained health detection model includes: Adding random noise to the first sample normal index feature to obtain a first updated sample normal index feature; Adding random noise to the first sample abnormal index feature to obtain a first updated sample abnormal index feature; Input the normal index features of the first updated sample into the health detection model to be trained, and obtain a first output result through encoding by the encoding network and decoding by the decoding network; Input the abnormal index features of the first updated sample into the health detection model to be trained, and obtain a second output result through encoding by the encoding network and decoding by the decoding network; Based on the third difference value between the first output result and the normal index features of the first sample, and the fourth difference value between the second output result and the abnormal index features of the first sample, adjust the parameters of the health detection model to be trained to obtain a preliminarily trained health detection model; Based on the preliminarily trained health detection model, obtain a pre-trained health detection model.

5. The method according to claim 4, wherein The obtaining the pre-trained health detection model based on the preliminarily trained health detection model includes: Add random noise to the normal index features of the second sample to obtain the second updated normal index features; Add random noise to the abnormal index features of the second sample to obtain the second updated abnormal index features; Input the second updated normal index features into the preliminarily trained health detection model, and obtain a first classification result through encoding by the encoding network and classification by the classification network; Input the second updated abnormal index features into the preliminarily trained health detection model, and obtain a second classification result through encoding by the encoding network and classification by the classification network; Based on the first classification result and the normal label, and the second classification result and the abnormal index, adjust the parameters of the preliminarily trained health detection model to obtain a pre-trained health detection model.

6. The method according to any one of claims 1 to 3, characterized in that The third historical index feature in the historical index feature set corresponds to a normal label, and the normal label is used to represent that the server cluster is normal. The fourth historical index feature in the historical index feature set corresponds to an abnormal label, and the abnormal label is used to represent that the server cluster is abnormal; The training the pre-trained health detection model by using the historical normal index features and historical abnormal index features of the server cluster to obtain a target health detection model includes: Add random noise to the third historical index feature to obtain the first updated historical index feature; Add random noise to the fourth historical index feature to obtain the second updated historical index feature; Input the first updated historical index feature into the pre-trained health detection model, obtain a third output result of the decoding network, and obtain a third classification result of the classification network; Input the second updated historical index feature into the pre-trained health detection model, obtain a fourth output result of the decoding network, and obtain a fourth classification result of the classification network; Based on the fifth difference value between the third output result and the third historical index feature, the sixth difference value between the fourth output result and the fourth historical index feature, the third classification result, the normal label, the fourth classification result, and the abnormal label, adjust the parameters of the pre-trained health detection model to obtain a target health detection model.

7. A server cluster detection device, characterized in that, The device includes: A generation model obtaining module, configured to train an index generation model to be trained by using a historical index feature set of a server cluster to obtain a normal index generation model, and train the index generation model by using the historical index feature set to obtain an abnormal index generation model; An index feature generation module, configured to generate sample normal index features of the server cluster through the normal index generation model, and generate sample abnormal index features of the server cluster through the abnormal index generation model; A first training module, configured to train a health detection model to be trained by using the sample normal index features and the sample abnormal index features to obtain a pre-trained health detection model; A second training module, configured to train the pre-trained health detection model by using the historical normal index features and the historical abnormal index features of the server cluster to obtain a target health detection model, where the target health detection model is used to detect the health of the server cluster.

8. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images