Fault prediction

Abstract

The present disclosure relates to a fault prediction method, an electronic device, a computer program product, and a storage medium. The method comprises: acquiring at least two types of data corresponding to the operation state of a target object; inputting the at least two types of data into a fault prediction model, wherein the fault prediction model generates, on the basis of feature cross processing, target cross features corresponding to the at least two types of data, and performs fault prediction on the basis of the target cross features; and acquiring fault prediction information output by the fault prediction model, wherein the fault prediction information is used for indicating whether a fault is about to occur in the target object.

Description

[0001] Fault prediction related applications

[0002]

[0001] This disclosure claims priority to Chinese patent application No. 202410367184.9, filed on March 27, 2024, entitled “Fault Prediction Method, Electronic Device, Computer Program Product, and Storage Medium,” the entire text of which is incorporated herein by reference.

[0003]

[0002] The present disclosure relates to the field of computer technology, and more particularly to fault prediction.

[0004] With the development of internet technology, more and more internet-based computing service models have emerged. Furthermore, various internet vendors provide users with corresponding computing services by building servers. These servers can be used to provide services to users, including, but not limited to, cloud computing platforms, online gaming services, and database services. Taking cloud computing platforms as an example, NC (Node Controller) downtime, resource exhaustion, and abnormal status are among the most common problems encountered in cloud computing platforms. These problems can severely impact the stability and reliability of cloud computing platforms. Furthermore, NC downtime, resource exhaustion, and abnormal status can impact users, such as inability to access data, inability to use services, and high access latency. In particular, prolonged NC downtime can even lead to serious consequences such as data loss and service interruptions. Furthermore, these failures can damage the platform's reputation and credibility, affecting users' trust in and willingness to use the platform.

[0005]

[0004] In related art, fault prediction technology is used to identify potential server faults in advance. This technology primarily uses monitoring data such as single indicators and performance parameters of the monitoring system to identify potential faults in advance, but is not applicable to diverse monitoring data. Therefore, the diversity of data poses challenges to fault prediction technology.

[0006]

[0005] The present disclosure provides a fault prediction method, an electronic device, a computer program product, and a storage medium to address deficiencies in related technologies.

[0007]

[0006] According to a first aspect of an embodiment of the present disclosure, a fault prediction method is proposed, comprising: obtaining at least two types of data corresponding to an operating state of a target object; inputting the at least two types of data into a fault prediction model, wherein the fault prediction model generates target cross-features corresponding to the at least two types of data based on feature cross-processing, and performs fault prediction based on the target cross-features; and obtaining fault prediction information output by the fault prediction model, wherein the fault prediction information is used to indicate whether a fault is about to occur in the target object.

[0008]

[0007] According to a second aspect of an embodiment of the present disclosure, a fault prediction method is also provided, comprising: obtaining at least two types of data corresponding to the operating status of a cloud server in a cloud computing platform; inputting the at least two types of data into a fault prediction model, and having the fault prediction model perform feature cross-processing on the at least two types of data to generate target cross-features; and determining, by the fault prediction model, fault prediction information based on the target cross-features, wherein the fault prediction information is used to indicate whether the cloud server is about to fail.

[0009]

[0008] According to a third aspect of an embodiment of the present disclosure, an electronic device is further provided, comprising a processor; and a memory for storing processor-executable instructions; wherein the processor implements the method described in the first aspect above by running the executable instructions.

[0010]

[0009] According to a fourth aspect of an embodiment of the present disclosure, a computer program product is provided, comprising a computer program, wherein when the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0011]

[0010] According to a fifth aspect of the embodiments of the present disclosure, a computer-readable storage medium is further provided, on which a computer program is stored. When the program is executed by a processor, the steps in the method described in the first aspect are implemented.

[0012]

[0011] The technical solutions provided by the embodiments of the present disclosure may include the following beneficial effects: As can be seen from the above embodiments, the present disclosure generates target cross-features corresponding to the at least two types of data based on feature cross-processing through a fault prediction model, and performs fault prediction based on the target cross-features. The fault prediction information outputted is generated by performing feature cross-processing on at least two types of data, so that the target cross-features obtained can include joint features of different types of data. When the target cross-features include joint features, the potential associations between different types of data can be characterized by the target cross-features, thereby enabling the fault prediction information obtained based on the target cross-features to more accurately predict whether a target object is about to fail. Therefore, the present disclosure can use multiple different types of data for fault prediction and can thereby effectively improve the accuracy of fault prediction.

[0013]

[0012] It should be understood that the above general description and the detailed description below are only exemplary and explanatory and cannot limit the present disclosure.

[0014]

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0015]

[0014] FIG1 is a schematic diagram of the architecture of a fault prediction system provided by an exemplary embodiment.

[0016]

[0015] FIG2 is a schematic flow chart of a fault prediction method according to an embodiment of the present disclosure.

[0017]

[0016] FIG3 is a schematic diagram of the architecture of a fault prediction model according to an embodiment of the present disclosure.

[0018]

[0017] FIG4 is a schematic diagram of the architecture of a fault prediction model according to another embodiment of the present disclosure.

[0019]

[0018] Figure 5 is a schematic flowchart of a fault prediction method applied to a cloud computing platform according to an embodiment of the present disclosure.

[0020]

[0019] Figure 6 is a structural diagram of an electronic device according to an embodiment of the present disclosure.

[0021]

[0020] FIG7 is a block diagram of a fault prediction device provided by an exemplary embodiment.

[0022]

[0021] FIG8 is a block diagram of a fault prediction device applied to a cloud computing platform provided by an exemplary embodiment.

[0023]

[0022] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. When the following description refers to the drawings, identical numerals in different drawings represent identical or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with one or more embodiments of the present disclosure. Rather, they are merely examples of apparatuses and methods consistent with certain aspects of one or more embodiments of the present disclosure, as detailed in the appended claims.

[0024] It should be noted that in other embodiments, the steps of the corresponding method are not necessarily performed in the order shown and described in this disclosure. In some other embodiments, the method may include more or fewer steps than those described in this disclosure. In addition, a single step described in this disclosure may be decomposed into multiple steps for description in other embodiments; and multiple steps described in this disclosure may be combined into a single step for description in other embodiments.

[0025]

[0024] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties. The collection, use, and processing of relevant data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0025] With the development of Internet technology, Internet services are becoming more and more numerous and complex. Therefore, the computing services provided by servers are becoming increasingly complex, and the occurrence of failures is inevitable. The above-mentioned servers may include, but are not limited to: cloud computing platforms, online game servers, database servers, etc.

[0026] A cloud computing platform is a computing service model based on internet technology that integrates and manages computing, storage, and network resources through cloud computing to provide various cloud services, such as cloud storage, cloud computing, and cloud security. NC downtime is one of the most common issues in cloud computing platforms, severely impacting their stability and reliability. NC downtime can impact users of the cloud computing platform, such as being unable to access data or use services. Prolonged NC downtime can even lead to serious consequences such as data loss and service interruptions. Furthermore, NC downtime can damage the reputation and credibility of the cloud computing platform, impacting users' trust in and willingness to use the platform.

[0027] An online game server is a server used to host multiplayer online games. It provides the game's infrastructure and network support, allowing players to connect and participate in the multiplayer mode of the game via the Internet. Common problems include, but are not limited to, performance issues, network issues, and server crashes.

[0028] A database server is a server specifically designed to store, manage, and provide data access. In businesses and organizations, database servers typically play a crucial role, responsible for storing and managing various types of data, including but not limited to user information, transaction records, product information, and the like. Common problems include but are not limited to performance issues, data loss, database connection issues, concurrency conflicts, and the like.

[0029]

[0029] In view of the numerous types of server failures described above, related technologies have proposed using fault prediction technology to reduce the risk of related failures. The application of fault prediction technology in various types of servers can significantly improve server stability and reliability. Fault prediction technology primarily monitors various server indicators and performance parameters to identify potential failures in advance and take appropriate preventive measures to address them.

[0030]

[0030] In the related art, commonly used fault prediction technologies include two categories: one uses Bidirectional Long Short-Term Memory (Bidirectional Long Short-Term Memory) networks for fault prediction, and the other uses MING for fault prediction. Each of these is briefly introduced below.

[0031] 1. Bi-LSTM can use numerical sequence data such as CPU utilization and memory utilization for fault prediction. However, the use of Bi-LSTM has the following disadvantages: (1) It only uses numerical sequence data, while ignoring the information brought by semantic sequence data such as system exception logs. (2) Bi-LSTM performs fault prediction through a cyclic process, which consumes a lot of time and may cause the previously extracted features to be forgotten later, which is not conducive to feature extraction of long sequence data types.

[0032] 2. MING is a method that extracts features from temporal data using an LSTM model and spatial data using a random forest model. The extracted features are then concatenated and fed into a ranking model.

[0032] In the ranking model, the probability of a fault is ranked, and finally a relatively better threshold is selected to determine whether a fault is likely to occur. As can be seen from the above, MING extracts features of the time and space dimensions and directly splices them for prediction. Therefore, the disadvantages of using MING for fault prediction are: (1) MING is based on spatial data for fault prediction and is not suitable for multimodal sequence data scenarios. (2) MING uses two independent models, ignoring the joint features of different types of data at the time scale.

[0033]

[0033] Therefore, the present disclosure addresses at least one of the above-mentioned deficiencies in the related art and proposes a corresponding fault prediction solution, which will be described in detail below with reference to the accompanying drawings.

[0034] FIG1 is a schematic diagram of the architecture of a fault prediction system provided by an exemplary embodiment. As shown in FIG1 , the system may include a server 11, a network 12, and several electronic devices 13 (eg, a PC (Personal Computer), a server), etc.

[0035]

[0035] The server 11 may be an object requiring fault prediction, and the server 11 may be a physical server including an independent host, or the server 11 may be a virtual server carried by a host cluster.

[0036]

[0036] In the technical solution of one or more embodiments of the present disclosure, the server 11 may be deployed with relevant functions for implementing the fault prediction method in this embodiment. For example, a program for implementing fault prediction may be run to perform fault prediction on the server 11 itself.

[0037]

[0037] In addition, in the technical solutions of one or more embodiments of the present disclosure, the relevant program for implementing the fault prediction method of the present embodiment may also be deployed on the electronic device 13. The electronic device 13 can then interact with the server 11 to obtain relevant data on the server 11 for fault prediction, thereby enabling fault prediction for the server 11 to be implemented based on the relevant program and the relevant data. The interaction between the electronic device 13 and the server 11 may specifically rely on data transmission between the electronic device 13 and the server 11. The data transmission may be implemented via the network 12, which may include various types of wired or wireless networks, and the present disclosure is not limited thereto.

[0038]

[0038] Furthermore, during operation, the electronic device 13 may run a client-side program for implementing fault prediction to implement the fault prediction function. It should be noted that the fault prediction client application may be pre-installed on the electronic device 13 so that the client can be started and run on the electronic device. Of course, when an online "client" such as HTML5 technology is used, the client can be obtained and run without installing the corresponding application on the electronic device.

[0039]

[0039] FIG2 is a schematic flow chart of a fault prediction method according to an embodiment of the present disclosure. As shown in FIG2, the fault prediction method shown in this embodiment may include the following steps.

[0040]

[0040] Step S202: Acquire at least two types of data corresponding to the operating status of the target object.

[0041] The fault prediction method provided in the embodiments of the present disclosure can be used in scenarios where fault prediction is performed on various types of servers, such as cloud computing platforms, online game servers, and database servers. In other words, the target objects can be servers for which fault prediction is required, including but not limited to the aforementioned cloud computing platforms, online game servers, and database servers.

[0042]

[0042] The at least two types of data may include, but are not limited to, the following types of data: log data, numerical data, image data, etc., which are not limited in this disclosure. Log data may include, for example, exception logs automatically recorded by the system, manually recorded exception logs, etc. Numerical data may include, but are not limited to, multiple time series data such as CPU utilization, memory utilization, and power consumption. Image data may include, for example, video data, one or more image data collected at a fixed period or according to a trigger condition, etc. The trigger condition may be visual information indicating a fault in the target object, such as the appearance of sparks or the entry of an illegal person into the computer room.

[0043] Of course, the specific types of data involved also depend on what kind of data can be generated in the scenario. In the embodiments of the present disclosure, processing is involved for at least two types of data. In fact, any type of data in the relevant art can be applied to the processing process of the embodiments of the present disclosure.

[0044]

[0044] Step S204: inputting at least two types of data into the fault prediction model, the fault prediction model generates target cross features corresponding to the at least two types of data based on feature cross processing, and performs fault prediction according to the target cross features.

[0045]

[0045] Specifically, after acquiring at least two types of data, the at least two types of data can be input into a fault prediction model. The fault prediction model can then process the at least two types of data using a feature cross-processing method to obtain target cross-features corresponding to the at least two types of data. After obtaining the target cross-features, fault prediction can be performed based on the target cross-features to obtain fault prediction information.

[0046]

[0046] Specifically, after acquiring at least two types of data, by performing feature crossover on the at least two types of data, the purpose of combining the features of the at least two types of data can be achieved. Thus, the feature crossover process can improve the accuracy and generalization ability of the model when performing fault prediction on multiple different types of data.

[0047]

[0047] In some optional embodiments, the fault prediction model includes: a data encoding module, a cross-attention module, and a fault prediction module; each module is introduced below.

[0048] The data encoding module is configured to map each type of data into a corresponding vector. The data encoding module can be configured to use different mapping methods for different types of data to map them into corresponding vectors. Since the input value data encoding module receives at least two types of data, a corresponding vector can be generated for each type of data.

[0049] In this embodiment, when the data is semantic data, the data encoding module can obtain a vector corresponding to the semantic data by mapping the elements in the semantic data. For example, the vector corresponding to the semantic data can be obtained by including but not limited to the following methods: using word embedding technology, such as Word2Vec (Word to Vector), GloVe

[0050] Global Vectors for Word Representation (GVWR) or FastText (FastText) map each word into a fixed-dimensional vector. These word vectors capture the semantic and grammatical relationships between words, thereby representing the meaning of natural language in computers. Furthermore, vectorized representation of semantics can be achieved through word-level pre-trained models (such as BERT (Bidirectional Encoder Representations from Transformers) and GPT (Generative Pre-trained Transformer). By mapping semantic data into vectors, these vectors can contain rich semantic information.

[0051] When the data is numerical data, it may include, but is not limited to, multiple time series data such as CPU utilization, memory utilization, and power consumption. For example, the data encoding module may map elements in the numerical data into vectors in a manner that includes, but is not limited to, one-hot encoding, word embeddings, feature extraction models, convolutional neural networks (CNNs), self-attention mechanisms, and the like.

[0052]

[0051] When the data is image data, the data encoding module may map elements in the image data into vectors in ways that include, but are not limited to, frame-level feature extraction, optical flow features, 3D convolutional neural networks, LSTM (Long Short-Term Memory) or GRU (Gated Recurrent Unit) models, video encoder (such as H.264, H.265, etc.) processing, attention mechanism, pre-trained model (for example, I3D (Interactive Three Dimensions), SlowFast (slow-fast architecture), etc.) processing, etc.

[0053]

[0052] For example, the disclosed embodiment can be processed using the fault prediction model shown in FIG3 . The data encoding module can further include a semantic sequence embedding module and a numerical sequence embedding module. Semantic sequence data (i.e., semantic data that can represent temporal relationships) from a target object can be input into the semantic sequence embedding module to map elements in the semantic sequence data into a first vector. Furthermore, numerical sequence data (i.e., numerical data that can represent temporal relationships) from the target object can be input into the numerical sequence embedding module to map elements in the numerical sequence data into a second vector using a convolution operation.

[0054] The cross-attention module is connected to the data encoding module and is used to perform feature cross-processing on the vector output by the data encoding module to obtain target cross-features. The cross-attention module is connected to the data encoding module, and the data encoding module can then input the generated vector into the cross-attention module. After obtaining the vector, the cross-attention module can perform feature cross-processing on the vector to obtain the target cross-feature.

[0055] Specifically, since the aforementioned vector generation is based on at least two types of data, when a corresponding vector is generated based on each type of data, at least two vectors can be obtained. Furthermore, feature crossover can be performed on the at least two vectors to combine the features of the at least two types of data and generate new features, i.e., obtain target crossover features. Feature crossover processing can improve the accuracy and generalization capability of the model. Optionally, feature crossover processing on the at least two vectors can be performed using methods including, but not limited to, the following: Factorization Machines (FM), Field-aware Factorization Machines (FFM), Deep Neural Networks (DNN), xDeepFM (Extreme Deep Factorization Machine), Cross Network

[0056] (Cross Network), Embedding Concatenation, Pooling Layers, Element-wise Product / Interaction Layer, Multi-Head Attention

[0057] (Multi-headed attention).

[0058] Furthermore, the specific method for performing feature cross-processing on multiple vectors needs to be selected based on the specific problem scenario and data characteristics. In practical applications, the results obtained by experimentally comparing different feature cross-processing methods can be used to find a more optimal feature cross-processing method, which will not be listed here.

[0059]

[0056] As an optional embodiment, the aforementioned steps are implemented by performing feature cross processing on the vector output by the data encoding module to obtain target cross features: for each vector, feature cross processing is performed on the vector and at least one other remaining vector to generate a cross feature corresponding to the vector; wherein the target cross feature includes each cross feature output by the cross attention module.

[0060]

[0057] In this embodiment, by inputting at least two vectors into the cross attention module, the cross attention module can perform feature cross processing on the at least two vectors, and then obtain cross features corresponding to each vector respectively after the feature cross processing of the at least two vectors. For example, when the at least two vectors include two vectors, two cross features can be obtained; further, the cross features corresponding to different vectors can be obtained after processing using the same feature cross method, but the cross features corresponding to different vectors can be features generated using different vectors obtained based on the at least two vectors. For example, when the cross attention module is implemented by the Cross Encoder in the Transformer architecture, and the at least two vectors include vector 1 and vector 2, vector 1 can be mapped to obtain vector 1a and vector 1b, and vector 2 can be mapped to obtain vector 2a and vector 2b by methods including but not limited to a fully connected layer, a shared weight matrix, and multi-head attention. Each of vector 1a, vector 1b, vector 2a, and vector 2b can include one or more sub-vectors, and then vector 1a, vector 1b, vector 2a, and vector 2b are mapped. Vector 2b is input into the same set of Cross Encoders in the Transformer architecture, so that Cross Encoder 1 and Cross Encoder? in this set of Cross Encoders can process vector 1 and vector 2 at the same time, and capture the interaction information between vector 1 and vector 2 through the attention mechanism, so that Cross Encoder 1 obtains the cross features corresponding to vector 1, and Cross Encoder 2 obtains the interaction features corresponding to vector 2. Furthermore, when the at least two vectors include vector 1, vector 2, and vector 3, combinations of every two vectors can be obtained, namely, vector 1 and vector 2, vector 2 and vector 3, and vector 1 and vector 3. Then, vector 1 and vector 2, vector 2 and vector 3, and vector 1 and vector 3 can be input into three sets of Cross Encoders, respectively, to obtain cross features 1a and 1b based on vector 1 and vector 2, cross features 2a and 2b based on vector 2 and vector 3, and cross features 3a and 3b based on vector 1 and vector 3. When the at least two vectors include four or more vectors, the cross feature corresponding to each vector can be determined with reference to the method in this embodiment, which will not be described in detail here.

[0061]

[0058] As an optional embodiment, when the cross-attention module includes a cross-attention submodule, a feature cross-processing is performed on the vector output by the data encoding module through the cross-attention submodule; when the cross-attention module includes multiple cross-attention submodules connected in sequence, multiple cross-attention submodules are used to iteratively perform multiple feature cross-processing on the vector output by the data encoding module. That is, the cross-attention module includes one or more cross-attention submodules connected in sequence. When the cross-attention module includes a cross-attention submodule, a feature cross-processing is performed on the vector output by the data encoding module through the cross-attention submodule; when the cross-attention module includes multiple cross-attention submodules, multiple cross-attention submodules are used to iteratively perform multiple feature cross-processing on the vector output by the data encoding module, and the data used by the latter cross-attention submodule to perform feature cross-processing is the result obtained after the former cross-attention submodule performs feature cross-processing. For example, as shown in FIG3 , a fault prediction model for performing fault prediction is provided. Taking sequence data as an example, the fault prediction model is used to predict faults for a target object based on two different types of sequence data. The cross-attention module in the fault prediction model includes two cross-attention sub-modules, and the two different types of data are semantic sequence data and numerical sequence data.

[0062]

[0059] In this embodiment, feature cross processing is performed on the first vector and the second vector of at least two vectors through a cross-attention module, including: performing cross-attention calculation on the first vector and the second vector through a cross-attention sub-module, and the last cross-attention sub-module outputs the first cross feature corresponding to the first vector and the second cross feature corresponding to the second vector. The process of implementing cross-attention calculation through any cross-attention submodule includes: inputting a first vector group into the first cross-encoder in any cross-attention submodule to obtain a first output result, and inputting a second vector group into the second cross-encoder in any cross-attention submodule to obtain a second output result; when any cross-attention submodule is located at the first position of all cross-attention submodules, the first vector group includes: a first query vector obtained based on the first vector, a second key vector obtained based on the second vector, and a second value vector; and the second vector group includes: a second query vector obtained based on the second vector, a first key vector obtained based on the first vector, and a first value vector; when the cross-attention submodule is not located at the first position of all cross-attention submodules, the first vector group includes: a query vector obtained based on the first output result of the previous cross-attention submodule, a key vector obtained based on the second output result of the previous cross-attention submodule, and a value vector; and the second vector group includes: a query vector obtained based on the second output result of the previous second cross-encoder, a key vector obtained based on the first output result of the previous first cross-encoder, and a value vector. As shown in Figure 3, in this embodiment, since fault prediction is based on two different types of sequence data, each cross-attention submodule includes two cross encoders. Cross encoder) to serve as the first cross encoder and the second cross encoder respectively.

[0063] In other words, the cross-attention module includes one or more layers of sequentially connected cross-attention sub-modules. Each layer of the cross-attention sub-module includes two cross-encoders, namely, a first cross-encoder and a second cross-encoder. Thus, a cross-attention calculation can be performed using the first cross-encoder and the second cross-encoder. In this embodiment, the first vector can be processed by the projection layer to generate the following: a first query vector, a first key vector, and a first value vector. The second vector can be processed by the projection layer to generate the following: a second query vector, a second key vector, and a second value vector.

[0064]

[0061] When the cross-attention sub-module a is the first of all cross-attention sub-modules included in the cross-attention module, in order to realize feature crossover through the first cross-attention sub-module, in this embodiment, the query vector, key vector and value vector corresponding to the first vector and the second vector are cross-processed, that is, a first vector group P1 including the first query vector, the second key vector and the second value vector is obtained, and a second vector group P2 including the second query vector, the first key vector and the first value vector is obtained, and then the first vector group P1 and the second vector group P2 are respectively input into the first cross encoder al and the second cross encoder a2 to obtain the first output result of the first cross encoder al and the second output result of the second cross encoder a2. Still taking Figure 3 as an example, the first cross-attention submodule a (Cross Encoder group) performs the following operations: In the cross-attention submodule a, the first vector and the second vector are simultaneously input and the cross attention of the two vectors is calculated. In the Cross Encoder based on Semantic Embedding (i.e., the first cross encoder al), the input Semantic Embedding generates the query (Q_S) (i.e., the first query vector), and the input Numeric Embedding generates the key (K_N).

[0065] (i.e., the second key vector) and value ( V_N ) (i.e., the second value vector), and then the query ( Q_S ) and key ( K_N ) are used to calculate the attention score 1 (i.e., the attention score), and the attention score 1 is combined with the value ( V_N ) to calculate the output of the Cross Encoder based on the Semantic Embedding, that is, the first output result RL is obtained. In the Cross Encoder based on Numeric Embedding (i.e., the second cross encoder a2 ), the input Numeric Embedding generates the query ( Q_N ) (i.e., the second query vector), and the input Semantic Embedding generates the key ( K_S ) (i.e., the first key vector) and value ( V_S ) (i.e., the first value vector), and then the query ( Q_N ) and key ( K_S ) are used to calculate the attention score 2, and the attention score 2 is combined with the value ( V_S ) to calculate the output of the Cross Encoder based on the Numeric Embedding. The output of the Cross Encoder based on Embedding is the second output result R2.

[0066]

[0062] When the cross-attention submodule b is not the first among all the cross-attention submodules included in the cross-attention module, in order to implement feature crossover through this non-first cross-attention submodule, in this embodiment, a first vector group Q1 is obtained, which includes a query vector obtained by the first output result of the first cross-encoder in the previous cross-attention submodule, a key vector obtained based on the second output result of the second cross-encoder in the previous cross-attention submodule, and a value vector; and a second vector group Q2 is obtained, which includes a query vector obtained by the second output result of the previous second cross-encoder, a key vector obtained based on the first output result of the previous first cross-encoder, and a value vector. Then, the first vector group Q1 and the second vector group Q2 are input into the first cross-encoder and the second cross-encoder, respectively, to obtain the first output result of the first cross-encoder and the second output result of the second cross-encoder. The previous cross-attention submodule is the cross-attention submodule located in the previous layer of the current cross-attention submodule. For example, as shown in FIG3 , when the cross-attention submodule b is located in the back layer of the cross-attention submodule a, the cross-attention submodule b is further processed based on the first output result R1 and the second output result R2 output by the previous cross-attention module a, and the first cross-encoder bl in the cross-attention submodule b inputs the query (Q_S,) obtained based on the first output result R1, and the key (K_N') and value (V_N') obtained based on the second output result R2, and the second cross-encoder b2 in the cross-attention submodule b inputs the query (Q_N) obtained based on the second output result R2. 5 ), and the key (K_S,) and value (V_S,) obtained based on the second output result R2.

[0067] When the first cross encoder and the second cross encoder are located in the last cross attention submodule, the first output result generated by the first cross encoder is used as the first cross feature, and the second output result generated by the second cross encoder is used as the second cross feature. Furthermore, the cross attention submodule can be a Transformer-based model.

[0068] A fault prediction module is connected to the cross-attention module and is configured to predict fault prediction information based on the target cross-features. The fault prediction module is connected to the cross-attention module to obtain the target cross-features output by the cross-attention module. After obtaining the target cross-features, a classification feature vector for classification can be obtained based on the target cross-features.

[0069]

[0065] In this embodiment, after the target cross-feature is determined, the target cross-feature can be linearly transformed and nonlinearly activated through, for example, a pooling layer (i.e., Pooler Layer) to obtain a classification feature vector, so as to extract key features from the target cross-feature, reduce overfitting, accelerate calculation, and improve the robustness and generalization ability of the model, thereby achieving the purpose of improving the performance and efficiency of the model.

[0070]

[0066] The classification feature vector may be a feature vector used to predict whether a target object will fail.

[0067] In this embodiment, fault prediction information indicating whether a target object will fail is obtained based on the classification feature vector. After the classification feature vector is determined, the classification feature vector may be input into a classifier for classification, so that the classifier can perform a classification operation based on the classification feature vector. In other words, the classification feature vector used by the classifier for classification may only include feature information contained in the target cross-feature.

[0071]

[0068] As an optional embodiment, the fault prediction model further includes: a parallel attention module; the parallel attention module is configured to perform feature extraction on each vector output by the data encoding module to obtain a parallel feature corresponding to each vector; wherein the fault prediction module is further connected to the parallel attention module and configured to perform fault prediction based on the target cross-feature and the parallel feature to obtain fault prediction information. In other words, feature extraction is performed on each vector to obtain a parallel feature corresponding to each vector, and feature extraction operations for each vector can be performed simultaneously. For example, when there are two vectors, namely a first vector and a second vector, feature extraction can be performed on the first vector to obtain a first parallel feature, and feature extraction can be performed on the second vector to obtain a second parallel feature. Optionally, feature extraction can be performed simultaneously by performing attention calculation on the first vector and the second vector to obtain a self-attention method to obtain the first parallel feature and the second parallel feature.

[0072]

[0069] As an optional embodiment, the parallel attention module includes a submodule group corresponding to each vector. When the submodule group includes one parallel attention submodule, feature extraction includes: performing an attention calculation on the corresponding vector once by the parallel attention submodule, and obtaining parallel features corresponding to the corresponding vector based on the self-attention extraction obtained by the calculation. When the submodule group includes multiple parallel attention submodules connected in sequence, feature extraction includes: performing multiple attention calculations on the corresponding vector multiple times by the multiple parallel attention submodules, and obtaining parallel features corresponding to the corresponding vector based on the self-attention extraction obtained by the calculation. Specifically, for any vector, the submodule group corresponding to the vector is used to perform attention calculation on the vector. Furthermore, generally, the number of parallel attention submodules corresponding to each vector is the same. When the number is greater than or equal to 2, the parallel attention submodules within the same submodule group are linearly connected in sequence. The input of the first parallel attention submodule in each submodule group is one of the at least two vectors, and the input vectors in different submodule groups are different. The inputs of the parallel attention submodules other than the first parallel attention submodule in each submodule group are derived based on the output of the previous parallel attention submodule. As shown in FIG4 , another fault prediction model for fault prediction is provided. In addition to the model shown in FIG3 , this fault prediction model also includes the parallel attention module described in the aforementioned embodiment, for performing fault prediction on a target object based on two different types of sequence data. The parallel attention module includes two submodule groups, submodule group A and submodule group B, each of which includes two sequentially connected parallel attention submodules.

[0073]

[0070] For example, taking the obtained vector including a first vector and a second vector as an example, attention calculation is performed on the first vector by one or more sequentially connected first parallel attention submodules to obtain a first self-attention. The first parallel attention submodule is a parallel attention submodule included in the submodule group corresponding to the first vector. Based on the first self-attention, a first parallel feature of the first vector is extracted. Furthermore, attention calculation is performed on the second vector by one or more sequentially connected second parallel attention submodules to obtain a second self-attention. The second parallel attention submodule is a parallel attention submodule included in the submodule group corresponding to the second vector. Based on the second self-attention, a second parallel feature of the second vector is extracted. Taking the example of multiple first parallel attention submodules performing attention calculation on a first vector, the first vector is fed into the first of the multiple first parallel attention submodules for self-attention calculation. The self-attention output by the first first parallel attention submodule is obtained. This self-attention is then fed into the second first parallel attention submodule for self-attention calculation. This cycle continues until the first self-attention output by the last of the multiple first parallel attention submodules is obtained. Feature extraction is then performed based on this first self-attention to obtain a first parallel feature corresponding to the first vector. The implementation method for obtaining parallel features corresponding to any vector can be referred to the implementation method for obtaining first parallel features and will not be further described here. Optionally, the parallel attention module can be a Transformer-based model. For example, as shown in FIG4 , when each sub-module group includes two parallel attention sub-modules, parallel attention calculation is performed through the parallel attention sub-module (Parallel Encoder, including Transformer): the query (Q_S), key (K_S), and value (V_S) obtained based on the first vector are input into the first parallel attention sub-module Al, and the query (Q_N), key (K_N), and value (V_N) obtained based on the second vector are input into the first parallel attention sub-module Bl, and attention (i.e., attention) calculation is performed to obtain self-attention, and then parallel features are extracted based on the self-attention.When the parallel attention sub-module is provided with multiple layers, it can be repeated multiple times. When the parallel attention sub-module is provided with two layers as shown in FIG4 , the first parallel attention sub-module A2 located in the second layer inputs the query (Q_S''), key (K_S''), and value (V_S'') determined based on the result obtained by the first parallel attention sub-module A1, and the second parallel attention sub-module B2 located in the second layer inputs the query (Q_N''), key (K_N''), and value (V_N'') determined based on the result obtained by the second parallel attention sub-module B1. That is, each sub-module group can include multiple parallel attention sub-modules.

[0074]

[0071] As an optional embodiment, after obtaining the parallel features corresponding to each vector, the fault prediction module can obtain a classification feature vector based on the target cross features and the parallel features. Furthermore, the fault prediction module can implement fault prediction based on the target cross features and the parallel features through the following steps to obtain fault prediction information: splicing the target cross features and the parallel features to obtain splicing features; performing linear transformation on the splicing features through the target linear layer in the fault prediction model to obtain a linear transformation result; performing nonlinear activation on the linear transformation result through the activation function in the fault prediction model to obtain a classification feature vector.

[0075] The target cross-feature in this embodiment can be generated in the manner described in the corresponding embodiment of FIG. 3 , and will not be described in detail here. After obtaining the parallel features and the target cross-feature, all of the parallel features and the target cross-feature can be concatenated to obtain a concatenated feature. Alternatively, each parallel feature can be concatenated to obtain a target parallel feature, and the cross-features corresponding to each vector can be concatenated to obtain a target cross-feature. Finally, the target parallel feature and the target cross-feature can be concatenated again to obtain a concatenated feature. Generally, both the parallel and cross-features are one-dimensional feature vectors; therefore, the concatenated feature obtained by concatenation is also a one-dimensional feature vector.

[0076] In this embodiment, after obtaining the splicing features, the splicing features can be linearly transformed by performing a linear operation on the target linear layer in the fault prediction model to obtain a linear transformation result corresponding to the splicing features. After obtaining the linear transformation result, the linear transformation result can be nonlinearly activated using an activation function in the fault prediction model (for example, a tanh() activation function can be used). This can introduce nonlinearity, solve the linear inseparability problem, increase the model's expressiveness, and suppress overfitting, thereby improving the performance and expressiveness of the fault prediction model.

[0077]

[0074] For example, taking fault prediction for a cloud computing platform as an example, after obtaining parallel features and obtaining target cross features according to the method in the above embodiment, as shown in FIG4 , the Pooler layer (i.e., the pooling layer) and the classifier are combined.

[0078] (Classifier) ​​obtains the classification feature vector: The Pooler layer performs a linear transformation on the concatenated features through the linear layer to obtain a linear transformation result, and uses the tanh() activation function to perform nonlinear activation on the linear transformation result to obtain the final classification feature vector used for classification.

[0079] After determining a classification feature vector based on the target cross-feature and the parallel feature, the classification feature vector can be input into a classifier for classification, so that the classifier can perform a classification operation based on the classification feature vector. In other words, the classification feature vector based on which the classifier performs classification can include feature information included in both the target cross-feature and the parallel feature.

[0080]

[0076] Optionally, when training a model implementing the disclosed method, the classifier output and sample labels are used to calculate cross-beam loss, and back propagation is performed, with the model parameters updated using the gradient descent method. Furthermore, the method for calculating cross-beam loss can be selected based on whether the classifier is used for binary classification or multi-classification. When performing binary classification, the results can be: the target object will fail or the target object will not fail; when performing multi-classification, the results can be: the target object will not fail or the target object will fail A or B. Taking a binary classification classifier as an example, when testing the trained model, the model is used to calculate the failure confidence of the test sample. When the confidence of the classifier output exceeds a certain threshold (e.g., 0.6, 0.7, etc., which can be selected based on the actual application), the sample is judged as positive, indicating that the test object corresponding to the sample will fail. Otherwise, it indicates that the test object corresponding to the sample will not fail.

[0081]

[0077] For example, still using FIG4 as an example, after obtaining the classification feature vector, the classifier can process the classification feature vector to obtain fault prediction information. Specifically, after receiving the classification feature vector output by the Pooler, the classifier performs a classification operation on the classification feature vector to obtain fault prediction information. When training the original model, the classifier output and sample labels are used to calculate the cross-beam loss, and backpropagation is performed. The model parameters are updated using the gradient descent method to obtain the final fault prediction model. During testing, the model is used to calculate the fault confidence of the test sample. When the confidence exceeds a certain threshold, the sample is judged as positive, indicating that the NC generating the sample will fail.

[0082]

[0078] Step S206: Obtaining fault prediction information output by the fault prediction model, wherein the fault prediction information is used to indicate whether the target object is about to fail.

[0083]

[0079] After the fault prediction model predicts and obtains fault prediction information based on at least two types of input data, the fault prediction information can be obtained by obtaining its output, thereby determining whether the target object is about to fail.

[0084]

[0080] The method disclosed herein generates target cross-features corresponding to the at least two types of data based on feature cross-processing using a fault prediction model, and performs fault prediction based on the target cross-features. The output fault prediction information can include joint features of different types of data, as the target cross-features are generated by performing feature cross-processing on the at least two types of data. When the target cross-features include joint features, the target cross-features can characterize potential associations between different types of data, thereby enabling the fault prediction information obtained based on the target cross-features to more accurately predict whether a target object is about to fail. Therefore, the present disclosure can use multiple different types of data for fault prediction, thereby effectively improving the accuracy of fault prediction. This effectively overcomes the technical problem in related arts of being unable to accurately predict faults using multiple different types of data.

[0085]

[0081] As an optional embodiment, the fault prediction model is obtained after training the original model; the method further includes: after each training of the original model, performing the following adjustment operation on the weight parameters of the first cross encoder and the second cross encoder in the same cross attention submodule: after each training of the original model, performing the following adjustment operation on the weight parameters of multiple cross encoders in the same cross attention submodule: adjusting the weight parameters of any cross encoder to the weight parameters of the remaining cross encoders. Taking the existence of two cross encoders as an example, adjusting the weight parameters of any cross encoder of the first cross encoder and the second cross encoder to the weight parameters of the other cross encoder. That is, a fault prediction model can be obtained by training the original model. Furthermore, after each training of the original model, the weight parameters of the first and second cross-encoders in the same layer of the original model (i.e., within the same cross-attention submodule) generally change. Therefore, the weight parameters of the first and second cross-encoders in the same layer generally differ. In this case, the present disclosure adjusts the weight parameters of either the first or second cross-encoder to the weight parameters of the other cross-encoder, thereby maintaining consistency between the weight parameters of the first and second cross-encoders in the same layer. For example, the weight parameters of the first cross-encoder can be modified to the weight parameters of the second cross-encoder, or the weight parameters of the second cross-encoder can be modified to the weight parameters of the first cross-encoder. Experimental measurements have shown that using the method of this embodiment to adjust the weight parameters of the first and second cross-encoders in the same cross-attention submodule can effectively improve the training efficiency and recognition accuracy of the original model. When there are three or more cross-encoders, the weight parameters of the cross-encoders can be adjusted in the same manner as described above, which will not be further described here.

[0086]

[0082] Based on the presence of a parallel attention module, the step of obtaining a classification feature vector for classification based on the target cross-feature includes the following steps: obtaining a classification feature vector based on the target cross-feature and the parallel feature. After performing feature extraction in different ways based on at least two vectors, the target cross-feature and the parallel feature can be obtained. To form a richer and more comprehensive feature representation, in this embodiment, the classification feature vector is obtained based on the target cross-feature and the parallel feature. In addition, obtaining the classification feature vector in this manner can provide more input information for the model, thereby improving the performance of the model, solving the problem of sparsity, and improving the robustness of the model.

[0087]

[0083] 1. Compared with Bi-LSTM, the fault prediction model implemented in the above prediction method has two major advantages: (1) In fault prediction, the system abnormality log information is added, which provides a more comprehensive understanding of the health status of the cloud computing system, thereby improving the accuracy of fault prediction. (2) Using a multimodal data fusion framework based on Transformer (i.e., Parallel Encoder and Cross Encoder both contain Transformer), it can effectively reduce the time of model training and better learn the long-term dependencies of long sequence data (i.e., semantic sequence data such as system abnormality logs).

[0088]

[0084] 2. Compared with MING, the fault prediction model that implements the above prediction method also has two major advantages: (1) Since MING is a prediction model specifically for spatial data and is not suitable for sequence data, the multimodal feature fusion technology based on Transformer is more suitable for feature extraction of sequence data. (2) It can learn the joint features of different modal data on a time scale, enhance the expression ability of different modal features, and improve the accuracy of fault prediction.

[0089]

[0085] FIG5 is a schematic flow chart of a fault prediction method applied to a cloud computing platform according to an embodiment of the present disclosure. As shown in FIG5, the fault prediction method applied to a cloud computing platform shown in this embodiment may include the following steps.

[0090]

[0086] Step S502: obtaining at least two types of data corresponding to the operating status of the cloud server in the cloud computing platform.

[0091]

[0087] The fault prediction method provided in the embodiment of the present disclosure can be used in scenarios where fault prediction is performed on cloud servers in a cloud computing platform.

[0092]

[0088] As an optional embodiment, the data types include semantic data, numerical data, and image data. Semantic data includes system logs of the cloud computing system, numerical data includes resource utilization and power consumption of the cloud computing system, and image data includes images captured from hosts in the cloud computing system. Log data may include exception logs automatically recorded by the system, manually recorded exception logs, etc. Numerical data may include, but is not limited to, multiple time series data such as CPU utilization, memory utilization, and power consumption. Image data may include video data, one or more image data captured at a fixed period or based on trigger conditions, etc. The trigger condition may be visual information indicating a fault in the target object, such as the appearance of sparks or illegal entry into the computer room.

[0093]

[0089] Step S504: input at least two types of data into the fault prediction model, and the fault prediction model performs feature cross processing on the at least two types of data to generate target cross features.

[0094]

[0090] In some optional embodiments, the fault prediction model includes: a data encoding module, a cross-attention module and a fault prediction module; the data encoding module is used to map each type of data into a corresponding vector.

[0095]

[0091] The W encoding module can be used to adopt different mapping methods for different types of data and map them into corresponding vectors. Since the input value data encoding module is at least two types of data, a corresponding vector can be generated for each type of data.

[0096]

[0092] In some optional embodiments, the vector output by the data encoding module is subjected to feature cross processing to obtain target cross features, including: for each vector, the vector is subjected to feature cross processing with at least one other remaining vector to generate a cross feature corresponding to the vector; wherein the target cross feature includes each cross feature output by the cross attention module.

[0097]

[0093] In some optional embodiments, when the cross-attention module includes a cross-attention sub-module, the vector output by the data encoding module is subjected to a feature cross-processing operation once through the cross-attention sub-module; when the cross-attention module includes a plurality of cross-attention sub-modules connected in sequence, the vector output by the data encoding module is subjected to a plurality of iterative feature cross-processing operations multiple times through the plurality of cross-attention sub-modules.

[0098]

[0094] The fault prediction model is obtained after training the original model; the method also includes: after each training of the original model, performing the following adjustment operation on the weight parameters of multiple cross encoders located in the same cross attention submodule: adjusting the weight parameters of any cross encoder to the weight parameters of the remaining cross encoders.

[0099]

[0095] The fault prediction model further includes: a parallel attention module; the parallel attention module is configured to perform feature extraction on each vector output by the data encoding module to obtain a parallel feature corresponding to each vector; wherein the fault prediction module is further connected to the parallel attention module and configured to perform fault prediction based on the target cross-features and the parallel features to obtain the fault prediction information.

[0096] In some optional embodiments, the parallel attention module includes a sub-module group corresponding to each vector; when the sub-module group includes a parallel attention sub-module, the feature extraction includes: performing an attention calculation on the corresponding vector once by the parallel attention sub-module, and obtaining a parallel feature corresponding to the corresponding vector based on the self-attention extraction obtained by the calculation; when the sub-module group includes multiple parallel attention sub-modules connected in sequence, the feature extraction includes: performing multiple attention calculations on the corresponding vector iteratively by the multiple parallel attention sub-modules, and obtaining a parallel feature corresponding to the corresponding vector based on the self-attention extraction obtained by the calculation.

[0100]

[0097] Step S506: The fault prediction model determines fault prediction information according to the target cross-features, wherein the fault prediction information is used to indicate whether the cloud server is about to fail.

[0101]

[0098] After the fault prediction model predicts fault prediction information based on at least two types of input data, the fault prediction information can be obtained by obtaining the output thereof. Furthermore, whether a cloud server in the cloud service platform is about to fail can be determined based on the fault prediction information.

[0102]

[0099] Specifically, the specific implementation of each step in this embodiment can refer to the specific implementation of the relevant embodiment of the fault prediction method shown in FIG2, and will not be repeated here.

[0103] FIG6 is a schematic block diagram of an electronic device provided by an exemplary embodiment. Referring to FIG6 , at the hardware level, the device includes a processor 602, an internal bus 604, a network interface 606, a memory 608, and a non-volatile memory 610. Of course, it may also include hardware required for other functions. One or more embodiments of the present disclosure may be implemented in software, such as by the processor 602 reading a corresponding computer program from the non-volatile memory 610 into the memory 608 and then executing it. Of course, in addition to software implementation, one or more embodiments of the present disclosure do not exclude other implementations, such as logic devices or a combination of software and hardware. In other words, the execution subject of the following processing flow is not limited to individual logic units, but may also be hardware or logic devices.

[0104]

[0101] The present disclosure also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of any of the above methods when executed by a processor.

[0105]

[0102] Referring to FIG. 7 , a fault prediction device may be applied to the device shown in FIG. 6 to implement the technical solution of the present disclosure. The fault prediction device may include:

[0106]

[0103] A first acquisition module 71 acquires at least two types of data corresponding to the operating status of the target object;

[0107]

[0104] A cross module 72 inputs the at least two types of data into a fault prediction model, and the fault prediction model generates target cross features corresponding to the at least two types of data based on feature cross processing, and performs fault prediction based on the target cross features;

[0108]

[0105] The second acquisition module 73 acquires the fault prediction information output by the fault prediction model, wherein the fault prediction information is used to indicate whether the target object is about to fail.

[0109]

[0106] As an optional embodiment, the fault prediction model includes: a data encoding module, a cross-attention module and a fault prediction module; the data encoding module is used to map each type of data into a corresponding vector; the cross-attention module is connected to the data encoding module and is used to perform feature cross processing on the vector output by the data encoding module to obtain target cross features; the fault prediction module is connected to the cross-attention module and is used to predict fault prediction information based on the target cross features.

[0110]

[0107] As an optional embodiment, the cross-attention module is configured to: for each vector, perform feature cross processing on the vector and at least one other remaining vector to generate a cross feature corresponding to the vector; wherein, the target cross feature includes each cross feature output by the cross-attention module.

[0111]

[0108] As an optional embodiment, the cross-attention module is configured as follows: when the cross-attention module includes a cross-attention sub-module, a feature cross-processing is performed on the vector output by the data encoding module through the cross-attention sub-module; when the cross-attention module includes multiple cross-attention sub-modules connected in sequence, multiple feature cross-processings are iteratively performed on the vector output by the data encoding module through multiple cross-attention sub-modules.

[0112]

[0109] As an optional embodiment, the fault prediction model is obtained after training the original model; it also includes an adjustment module, which is configured to: after each training of the original model, perform the following adjustment operation on the weight parameters of multiple cross encoders located in the same cross attention submodule: adjust the weight parameters of any cross encoder to the weight parameters of the remaining cross encoders.

[0113]

[0110] As an optional embodiment, the fault prediction model also includes: a parallel attention module; a parallel attention module, which is used to perform feature extraction on the vectors output by the data encoding module respectively to obtain parallel features corresponding to each vector; wherein the fault prediction module is also connected to the parallel attention module, which is used to perform fault prediction based on the target cross features and the parallel features to obtain fault prediction information.

[0114]

[0111] As an optional embodiment, the parallel attention module includes a sub-module group corresponding to each vector respectively; when the sub-module group includes a parallel attention sub-module, the feature extraction includes: performing an attention calculation on the corresponding vector once through the parallel attention sub-module, and obtaining the parallel features corresponding to the corresponding vector based on the self-attention extraction obtained by the calculation; when the sub-module group includes multiple parallel attention sub-modules connected in sequence, the feature extraction includes: iteratively performing multiple attention calculations on the corresponding vector through multiple parallel attention sub-modules, and obtaining the parallel features corresponding to the corresponding vector based on the self-attention extraction obtained by the calculation.

[0115]

[0112] Referring to FIG. 8 , a fault prediction device for a cloud computing platform can be applied to the device shown in FIG. 6 to implement the technical solution of the present disclosure. The fault prediction device can include:

[0113] a first acquisition module 81 for acquiring at least two types of data corresponding to the operating status of a cloud server in the cloud computing platform;

[0116]

[0114] A cross module 82 is used to input at least two types of data into the fault prediction model, and the fault prediction model performs feature cross processing on the at least two types of data to generate target cross features;

[0117]

[0115] The second acquisition module 83 is used for the fault prediction model to determine fault prediction information according to the target cross-feature, wherein the fault prediction information is used to indicate whether the cloud server is about to fail.

[0118]

[0116] As an optional embodiment, the fault prediction model includes: a data encoding module, a cross-attention module and a fault prediction module; the data encoding module is used to map each type of data into a corresponding vector; the cross-attention module is connected to the data encoding module and is used to perform feature cross processing on the vector output by the data encoding module to obtain target cross features; the fault prediction module is connected to the cross-attention module and is used to predict fault prediction information based on the target cross features.

[0119]

[0117] As an optional embodiment, the data types include: semantic data, numerical data and image data, wherein the semantic data includes the system log of the cloud computing system, the numerical data includes the resource utilization and power consumption of the cloud computing system, and the image data includes the image collected from the host of the K computing system.

[0120]

[0118] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer. The computer may be in the form of a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email transceiver, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0121]

[0119] In a typical configuration, a computer includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0122]

[0120] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash RAM. o Memory is an example of a computer-readable medium.

[0123]

[0121] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be used to store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage, quantum memory, graphene-based storage media or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media, such as modulated data signals and carrier waves.

[0124]

[0122] It should also be noted that the terms "comprise," "include," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or device comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, commodity, or device. In the absence of further limitations, an element defined by the phrase "comprising a . . . . " does not preclude the presence of other identical elements in the process, method, commodity, or device comprising the element.

[0125]

[0123] The above description describes specific embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0126]

[0124] The terms used in one or more embodiments of the present disclosure are for the purpose of describing specific embodiments only and are not intended to limit the one or more embodiments of the present disclosure. The singular forms "a", "an", "the" and "the" used in one or more embodiments of the present disclosure and the appended claims are also intended to include the plural forms unless the context clearly states otherwise. Any or all possible combinations of the listed items.

[0127]

[0125] It should be understood that although one or more embodiments of the present disclosure may use terms such as first, second, and third to describe various information, such information should not be limited to these terms. These terms are merely used to distinguish information of the same type from one another. For example, without departing from the scope of one or more embodiments of the present disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the term "if" as used herein may be interpreted as "at the time of," "when," or "in response to determining."

[0128]

[0126] The above description is only a preferred embodiment of one or more embodiments of the present disclosure and is not intended to limit one or more embodiments of the present disclosure. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of one or more embodiments of the present disclosure should be included in the scope of protection of one or more embodiments of the present disclosure.

Claims

Claims 1. A fault prediction method, comprising: Acquire at least two types of data corresponding to the operating status of the target object; Inputting the at least two types of data into a fault prediction model, wherein the fault prediction model: generates target cross features corresponding to the at least two types of data based on feature cross processing, and performs fault prediction according to the target cross features; Fault prediction information output by the fault prediction model is obtained, wherein the fault prediction information is used to indicate whether the target object is about to fail.

2. The method according to claim 1, wherein: The fault prediction model includes: a data encoding module, a cross-attention module and a fault prediction module; the data encoding module is used to map each type of data into a corresponding vector; the cross-attention module is connected to the data encoding module and is used to perform feature cross processing on the vector output by the data encoding module to obtain the target cross feature; the fault prediction module is connected to the cross-attention module and is used to predict the fault prediction information based on the target cross feature.

3. The method according to claim 2, wherein: The performing feature cross processing on the vector output by the data encoding module to obtain target cross features includes: for each vector, performing feature cross processing on the vector and at least one other remaining vector to generate a cross feature corresponding to the vector; wherein the target cross feature includes each cross feature output by the cross attention module.

4. The method according to claim 2, wherein: In the case where the cross-attention module includes a cross-attention sub-module, the vector output by the data encoding module is subjected to a feature cross-processing operation once through the cross-attention sub-module; in the case where the cross-attention module includes a plurality of cross-attention sub-modules connected in sequence, the vector output by the data encoding module is subjected to a plurality of iterative feature cross-processing operations multiple times through the plurality of cross-attention sub-modules.

5. The method according to claim 4, wherein: The fault prediction model is obtained after training the original model; the method also includes: after each training of the original model, performing the following adjustment operation on the weight parameters of multiple cross encoders located in the same cross attention sub-module: adjusting the weight parameters of any cross encoder to the weight parameters of the remaining cross encoders.

6. The method according to claim 2, wherein: The fault prediction model further includes: a parallel attention module; 22 The parallel attention module is used to extract features from the vectors output by the data encoding module respectively to obtain parallel features corresponding to each vector; wherein the fault prediction module is also connected to the parallel attention module to jointly perform fault prediction based on the target cross features and the parallel features to obtain the fault prediction information.

7. The method according to claim 6, wherein: The parallel attention module includes a sub-module group corresponding to each vector respectively; in the case where the sub-module group includes one parallel attention sub-module, the feature extraction includes: performing an attention calculation on the corresponding vector once by the one parallel attention sub-module, and extracting the parallel features corresponding to the corresponding vector based on the self-attention obtained by the calculation; in the case where the sub-module group includes multiple parallel attention sub-modules connected in sequence, the feature extraction includes: performing multiple attention calculations on the corresponding vector iteratively by the multiple parallel attention sub-modules, and extracting the parallel features corresponding to the corresponding vector based on the self-attention obtained by the calculation.

8. A fault prediction method, comprising: Acquiring at least two types of data corresponding to the operating status of a cloud server in a cloud computing platform; inputting the at least two types of data into a fault prediction model, and causing the fault prediction model to perform feature cross-processing on the at least two types of data to generate target cross-features; The fault prediction model determines fault prediction information based on the target cross-feature, wherein the fault prediction information is used to indicate whether the cloud server is about to fail.

9. The method according to claim 8, wherein: The fault prediction model includes: a data encoding module, a cross-attention module and a fault prediction module; the data encoding module is used to map each type of data into a corresponding vector; the cross-attention module is connected to the data encoding module and is used to perform feature cross processing on the vector output by the data encoding module to obtain the target cross feature; the fault prediction module is connected to the cross-attention module and is used to predict the fault prediction information based on the target cross feature.

10. The method according to claim 9, wherein: The data types include: semantic data, numerical data, and image data. The semantic data includes the system log of the cloud computing system, the numerical data includes the resource utilization and power consumption of the cloud computing system, and the image data includes images collected from the host of the cloud computing system.

11. An electronic device, comprising: a processor; A memory for storing processor-executable instructions; wherein, The processor is configured to implement the method of any one of claims 1 to 10.

12. A computer program product comprising a computer program, wherein: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

13. A computer-readable storage medium having a computer program stored thereon, wherein: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

Images