Computer system fault diagnosis method and system for multi-modal perception data

Abstract

The application discloses a kind of multi-modal perception data computer system fault diagnosis method and system, the method of the application includes the unified vectorization characterization of the multi-modal perception data of computer system, for the data of each modality in multi-modal perception data, respectively utilize the expert network of multiple channels to generate vectorization feature, according to the data modality of input, share routing network and the modality routing network of corresponding modality are cascaded to calculate the trust degree of the data of one modality to different expert networks, and the feature representation of this data modality is obtained based on trust degree to each expert network feature output weighted summation;The feature representation of each modality is characterized to realize superposition;The feature obtained after feature confusion is sent into classification network to obtain the fault diagnosis result of computer system.The application aims to comprehensively use multi-modal perception data to realize the fault diagnosis of unified computer system, to solve the inconsistency problem when different modal independent fault diagnosis.

Description

Technical Field

[0001] This invention relates to fault diagnosis technology in the field of large-scale computer system operation and maintenance, specifically to a computer system fault diagnosis method and system based on multimodal sensing data. Background Technology

[0002] Large-scale computer systems rely on operation and maintenance management systems (also known as monitoring systems) to perceive and collect operational status, enabling dynamic monitoring and evaluation of infrastructure and software services. Among these, fault diagnosis based on perceived data can automatically analyze the types of anomalies in the system's state, supporting operation and maintenance personnel to quickly locate the cause of the fault and guiding the system to carry out intelligent operation and maintenance operations such as fault self-healing, which is crucial for maintaining the reliable and stable operation of the system.

[0003] With the widespread adoption of cloud-native technologies, large-scale computer systems are increasingly using microservice architectures to deploy basic software and services, supporting flexible service updates and operational iterations. The use of microservice architectures significantly increases the number of objects to be monitored during system operation, transforming the monitoring of previously independent service instances into monitoring a larger-scale instance connection graph, thus significantly increasing the difficulty of monitoring operational status. Furthermore, because an anomaly in one service instance can propagate to other services through the call chain between instances, leading to anomalies in multiple services, the monitoring of microservice operations in large-scale computer systems faces multi-source anomalies, and the corresponding awareness data manifests in multiple modalities, including logs, metrics, and call data.

[0004] Existing fault diagnosis methods for large-scale computer systems mainly rely on the independent analysis of perception data from different modalities to assess the types of faults occurring in the system. Currently, the industry's single-modal fault diagnosis methods perform log text parsing, indicator evaluation, and call analysis in isolation. However, the perception data for faults often exhibits inconsistent behavior across different modalities, leading to ambiguity in fault diagnosis. (a) Log modal perception data contains semantic events in the system, but merely recording events cannot provide call relationships and system service status context, potentially leading to omissions or misjudgments; (b) Indicator modal perception data is unstable, prone to instantaneous fluctuations, and lacks system service context semantics. Relying solely on indicator data for fault classification will result in numerous misjudgments; (c) Call modal data provides coarse-grained descriptions of fault characteristics, failing to provide specific information within system service instances and thus unable to provide concrete evidence for diagnosis. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a computer system fault diagnosis method and system based on multimodal sensing data, which addresses the above-mentioned problems in the prior art. The present invention aims to comprehensively utilize multimodal sensing data to achieve unified computer system fault diagnosis, thereby solving the inconsistency problem of independent fault diagnosis of a single modality.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A method for diagnosing computer system faults using multimodal sensing data includes the following steps:

[0008] S101, unify the vectorization representation of the multimodal sensing data of the computer system, including: for each modality of the multimodal sensing data, generate vectorized features using expert networks of multiple channels respectively; cascade the shared routing network and the modal routing network of the corresponding modality according to the input data modality to calculate the trust degree of the sensing data of the data modality to different expert networks; and obtain the feature representation of the data modality by weighted summing of the feature outputs of each expert network based on the trust degree.

[0009] S102, feature obfuscation is performed on the feature representations of each modality to achieve superposition;

[0010] S103, The features obtained after feature confusion are fed into the classification network to obtain the fault diagnosis results of the computer system.

[0011] Optionally, the multimodal perception data of the computer system in step S101 includes three modalities: log modal data, indicator modal data, and call modal data. The log modal data is the log data of the computer system, which includes metadata server operation event records and object allocation and event records of the object storage server. The indicator modal data is the indicator data of the computer system, which includes server temperature and voltage, and hardware resource utilization in the operating system, including CPU utilization, memory utilization, disk utilization, and network utilization. The call modal data is the microservice call data of the computer system. The fault diagnosis results include some or all of the following: network failure, process crash, login failure, access denied, and file loss.

[0012] Optionally, in step S101, when generating vectorized features using expert networks with multiple channels, the functional expression of the vectorized features generated by the expert networks with multiple channels is:

[0013] ;

[0014] ;

[0015] ;

[0016] in, , and These are vectorized features generated by expert networks across multiple channels based on log modal data, indicator modal data, and call modal data. , and The first The system monitors log modal data, indicator modal data, and call modal data within a given time interval. ~ They are respectively number 1 to A network of experts across multiple channels. This represents the total number of expert networks.

[0017] Optionally, in step S101, when the shared routing network and the corresponding modal routing network are concatenated according to the input data modality to calculate the trust level of a modality's data to different expert networks, the corresponding modal routing network includes a log routing network. Indicator routing network The call chain is determined by the network. And log routing network Indicator routing network The call chain is determined by the network. Share a global cross-modal routing network for use as a shared routing network The global cross-modal routing network The global cross-modal routing network is switched on. The modality routing network corresponding to the input data modality is connected so that only the corresponding modality routing network is activated when aggregating the outputs of multiple expert networks for a given data modality. The expression for calculating the trust level of a data modality with different expert networks is as follows:

[0018] ;

[0019] ;

[0020] ;

[0021] in, , and These represent the trust levels of different expert networks based on log modal data, indicator modal data, and call modal data, respectively. , and These represent the log modal data, indicator modal data, and call modal data sensed during the j-th system monitoring time interval, respectively. These are the weighting coefficients for shared routes and modal routes. For a global cross-modal routing network, Routing network for logs, For the indicator routing network, The call link is provided by the network.

[0022] Optionally, the function expression for obtaining the feature representation of the data modality by weighted summation of the feature outputs of each expert network based on trust level in step S101 is as follows:

[0023] ;

[0024] ;

[0025] ;

[0026] in, , and These are feature representations for log modal data, metric modal data, and call modal data, respectively. , and These represent the log modal data, indicator modal data, and call modal data sensed during the j-th system monitoring time interval, respectively. , and These represent the trust levels of different expert networks based on log modal data, indicator modal data, and call modal data, respectively. Represents element-wise product. , and These are vectorized features generated by expert networks across multiple channels based on log modal data, indicator modal data, and call modal data.

[0027] Optionally, in step S102, when performing feature obfuscation on the feature representations of each modality to achieve superposition, the method includes first calculating the occurrence frequency of a modality's feature representation at the current time and its occurrence frequency in historical sensing data using the term frequency-inverse document frequency method; calculating the weight of each modality's feature representation based on its occurrence frequency at the current time and its occurrence frequency in historical sensing data; and then weighting and summing the feature representations of each modality according to the weights to achieve superposition of the feature representations of each modality.

[0028] Optionally, in step S103, when the features obtained after feature confusion are fed into the classification network to obtain the fault diagnosis results of the computer system, the classification network used is a three-layer perceptron.

[0029] Furthermore, the present invention also provides a computer system fault diagnosis system for multimodal sensing data, including a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to execute a computer system fault diagnosis method for the multimodal sensing data.

[0030] Compared with existing technologies, this invention mainly achieves the following beneficial effects: The widespread application of cloud-native technologies in large-scale computer systems leads to increasingly complex system operating states, specifically manifested in the multi-dimensional and heterogeneous multimodal nature of state-aware data. When diagnosing system faults and anomalies, independently analyzing different modalities of perception data can lead to diagnostic ambiguities and conflicts due to varying sensitivities to different faults, affecting or even misleading system state assessment and fault maintenance. This embodiment designs a unified vectorized representation process for data (multi-expert network representation based on multimodal routing) to calculate feature space-aligned perception data feature representations for different modalities of perception data. This enables feature aggregation (feature obfuscation) and fault classification of aggregated features (classification network) of multimodal perception data, eliminating inconsistencies in the diagnosis of system anomalies across different modalities. Specifically, this embodiment designs dedicated and shared vector representation routing modules (multimodal routing) for different modalities of data, allowing different expert network channels to represent different modalities differently, effectively preserving the different sensitivity attributes of different modalities to different faults. This invention achieves aggregated fault diagnosis by using a unified vectorization representation process for sensing data of different modalities and based on the aligned vector representation of the data of different modalities. It can comprehensively utilize multimodal sensing data to achieve unified system fault diagnosis of computer systems and solve the inconsistency problem of independent fault diagnosis of single modality. Attached Figure Description

[0031] Figure 1 This is a schematic diagram illustrating the basic principle of the method in an embodiment of the present invention. Detailed Implementation

[0032] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0033] like Figure 1 As shown, the computer system fault diagnosis method using multimodal sensing data in this embodiment includes the following steps:

[0034] S101, unifying the vectorization representation of the multimodal sensing data of the computer system, including: for each modality of the multimodal sensing data, generating vectorized features using expert networks of multiple channels respectively; cascading the shared routing network and the corresponding modality routing network according to the input data modality to calculate the trust degree of the sensing data of that data modality to different expert networks; and weighting and summing the feature outputs of each expert network based on the trust degree to obtain the feature representation of that data modality, thereby realizing the allocation of feature generation of different modal data to the feature representation of the data modality in different proportions;

[0035] S102, perform feature confusion on the feature representations of each modality to achieve superposition;

[0036] S103, send the features obtained after feature confusion into a classification network to obtain the fault diagnosis result of the computer system.

[0037] As Figure 1 shown, in step S101 of this embodiment, the multi-modal perception data of the computer system respectively includes three types of modal data: log modal data, metric modal data, and call modal data. Among them, the log modal data is the log data of the computer system, and the log data includes metadata server operation event records and object allocation and event records of the object storage server. The metric modal data is the metric data of the computer system, and the metric data includes the temperature and voltage of the server, and the hardware resource utilization rate in the operating system. The hardware resource utilization rate includes CPU utilization rate, memory utilization rate, disk utilization rate, and network utilization rate. The call modal data is the microservice call data of the computer system. For example, if microservice A calls microservice B, the corresponding microservice call data is a triple <A, B, call time>. Vectorized feature generation is performed on the same modal data through expert networks of multiple channels. The multi-modal routing module cascades the shared routing network and the routing network of the corresponding modality according to the input data modality, and calculates the trust degree of a piece of data for different expert networks. The feature representations of each expert network are weighted and summed based on the trust degree to obtain the feature representation of this data. Feature confusion superimposes the features of multiple modalities and sends them into the classification network to output the fault category diagnosis result. The fault diagnosis result includes network failure, process crash, login failure, access denied, and file loss.

[0038] The main functions of the method of this embodiment include multi-expert network representation, feature confusion, and classification network based on multi-modal routing, which are respectively used for unified representation of multi-modal perception data, representation aggregation of multi-modal perception data, and fault classification based on aggregated representation. As Figure 1 shown, in this embodiment, P (P≥2, and the quantity can be adjusted according to computing resource constraints) expert networks are used for feature extraction of perception data. By using multiple expert networks, it is avoided that one (or more) expert cannot resolve the fault mode of the input data, providing stronger generalization for the representation process. Specifically, multiple Transformer blocks are used as expert networks. The main consideration is that this model architecture can calculate all positions of the input data sequence simultaneously, efficiently carry out training and inference, and its self-attention mechanism allows each expert neural network to extract features sensitive to faults in the input data. Each expert network consists of N blocks (N≥1, and the number of blocks is adjustable) of Transformer blocks. Given the The system monitors three data points across three modalities: logs, metrics, and call chains, within a given monitoring time interval. , and The first in the network The expert's calculation outputs for the three modes are as follows: , Heti It can be abbreviated as: , Heti In step S101 of this embodiment, when generating vectorized features using expert networks with multiple channels, the functional expression of the vectorized features generated by the expert networks with multiple channels is as follows:

[0039] ;

[0040] ;

[0041] ;

[0042] in, , and These are vectorized features generated by expert networks across multiple channels based on log modal data, indicator modal data, and call modal data. , and These represent the log modal data, indicator modal data, and call modal data sensed during the j-th system monitoring time interval, respectively. ~ They are respectively number 1 to A network of experts across multiple channels. This represents the total number of expert networks.

[0043] A multimodal routing network is used to calculate the trust level of a perceptual data point to different expert networks, assigning higher weights to expert networks with higher trust levels, and aggregating the outputs of each expert network based on the weights to obtain a vector representation of the data point. Given that classic multi-expert networks use a single route to integrate expert outputs in each task, it is difficult to balance unimodal and cross-modal features in a multimodal environment. Therefore, this embodiment proposes a four-fold routing mechanism that maintains a separate routing network for each modality, including a log routing network. Indicator routing network The call chain is determined by the network. and share a global cross-modal routing network. A switch connects the shared routing network to the modality routing network corresponding to the input data modality. This means that when aggregating the outputs of multiple experts on a single dataset, only the corresponding modality routing network is activated. Taking log data routing calculation as an example, when calculating expert network trust, the shared routing network and the log routing network are connected. These two networks then evaluate the shared trust and the log modality trust, respectively.

[0044] ;

[0045] in, It is a trainable parameterized matrix of the routing network. and Both are P-dimensional trust matrices. By merging the trust calculation results of the two routing networks, the trust level of the log modality data among the P experts in the j-th time interval can be calculated as follows:

[0046] ;

[0047] in, Here are the weighting coefficients for shared routing and modal routing. Finally, the outputs of the P experts are weighted and summed using the trust level of the routing network output as the weight, resulting in the final representation of the log modal awareness data:

[0048] ;

[0049] in, This represents element-wise multiplication. Similarly, a representation of an index modal and a call chain modal can be calculated. and Therefore, in step S101 of this embodiment, when the shared routing network and the corresponding modal routing network are concatenated according to the input data modality to calculate the trust level of a modality's data with different expert networks, the corresponding modal routing network includes the log routing network. Indicator routing network The call chain is determined by the network. And log routing network Indicator routing network The call chain is determined by the network. Share a global cross-modal routing network for use as a shared routing network The global cross-modal routing network The global cross-modal routing network is switched on. The modality routing network corresponding to the input data modality is connected so that only the corresponding modality routing network is activated when aggregating the outputs of multiple expert networks for a given data modality. The expression for calculating the trust level of a data modality with different expert networks is as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] in, , and These represent the trust levels of different expert networks based on log modal data, indicator modal data, and call modal data, respectively. , and These represent the log modal data, indicator modal data, and call modal data sensed during the j-th system monitoring time interval, respectively. These are the weighting coefficients for shared routes and modal routes. For a global cross-modal routing network, Routing network for logs, For the indicator routing network, The call link is provided by the network.

[0054] Therefore, in step S101 of this embodiment, the function expression for obtaining the feature representation of the data modality by weighted summation of the feature outputs of each expert network based on the trust level is as follows:

[0055] ;

[0056] ;

[0057] ;

[0058] in, , and These are feature representations for log modal data, metric modal data, and call modal data, respectively. , and These represent the log modal data, indicator modal data, and call modal data sensed during the j-th system monitoring time interval, respectively. , and These represent the trust levels of different expert networks based on log modal data, indicator modal data, and call modal data, respectively. Represents element-wise product. , and These are vectorized features generated by expert networks across multiple channels based on log modal data, indicator modal data, and call modal data.

[0059] After multi-expert network representation based on multimodal routing, feature vectors aligned with feature dimensions are obtained for different modal sensing data. To classify all sensing data comprehensively, this embodiment proposes to further aggregate the feature vectors of each sensing data point within the same time period using a weighted summation method. Specifically, term frequency-inverse document frequency (TF-IDF) is used to calculate the occurrence frequency of a data point at the current moment and its frequency in historical sensing data. Data that occurs frequently in the current fault but is relatively rare in historical monitoring has a stronger indicative ability and sensitivity to the current fault state of the system, and is assigned a higher weight when its features are aggregated with other data features. Specifically, in step S102 of this embodiment, when feature obfuscating the feature representations of each modality to achieve superposition, the following steps are taken: first, the term frequency-inverse document frequency method is used to calculate the occurrence frequency of a modality's feature representation at the current moment and its occurrence frequency in historical sensing data; then, the weights of the feature representations of each modality are calculated based on the occurrence frequency at the current moment and its occurrence frequency in historical sensing data; finally, the feature representations of each modality are weighted and summed according to the weights to achieve superposition of the feature representations of each modality. It should be noted that Term Frequency-Inverse Document Frequency (TF-IDF) is a well-known statistical method used to evaluate the importance of a word to a document within a document set or corpus. The importance of a word increases proportionally to the number of times it appears in a document, but decreases inversely proportionally to its frequency in the corpus. In this embodiment, when using TF-IDF to calculate weights, for any data point of any modality (which could be a log, metric, or call), the frequency of its contained keywords in all historical data is denoted as f1, and the frequency of the keyword in the current window is denoted as f2. The importance of this data point is f2 / f1; that is, the less frequently the data appears in historical data and the more frequently it appears in the current window, the more important it is, and the higher its weight. In the current window, the f2 / f1 values ​​of multiple data points are normalized to the (0,1) interval, which represents the weight of the corresponding data point.

[0060] As an optional implementation, in step S103, when the features obtained after feature confusion are fed into the classification network to obtain the fault diagnosis result of the computer system, a three-layer perceptron is used as the classification network. Using a three-layer perceptron as the classification network, the input is the aggregated features of multiple modal data perceived in the same time period, and the output is the fault classification. The three-layer perceptron is a multilayer perceptron (MLP) with three layers, a fully connected neural network consisting of an input layer, hidden layers, and an output layer. Each layer is composed of neurons and is computed in a feedforward manner, meaning the output of each layer serves as the input of the next layer.

[0061] To verify the computer system fault diagnosis method using multimodal sensing data in this embodiment, tests were conducted on the General Intelligent Operation and Maintenance Dataset GAIA and the Railway Ticketing System Dataset TT, both running in real-world large data centers. The traditional log-based method Cloud19 was employed, along with this method that removes shared routing networks and modal routing networks (including log routing networks). Indicator routing network The call chain is determined by the network. For comparison, the simplified method of ) is used as a reference, and the final test results are shown in Table 1 below.

[0062] Table 1: Comparison of Test Results from Multiple Methods

[0063]

[0064] As shown in Table 1, on the General Intelligent Operation and Maintenance Dataset GAIA and the Railway Ticketing System Dataset TT, the computer system fault diagnosis method using multimodal sensing data in this embodiment (this method) provides better fault classification accuracy, recall, and F1 score compared to the traditional log-based method Cloud19 using single-modality data. Without using the computer system fault diagnosis method using multimodal sensing data in this embodiment (this method), the shared routing network and modal routing network (including log routing network) would be significantly better. Indicator routing network The call chain is determined by the network. If the fault classification effect is significantly reduced, it confirms the effectiveness of the shared routing network and modal routing network in the computer system fault diagnosis method for multimodal sensing data in this embodiment (this method).

[0065] In summary, in classic hybrid expert network models, when using multi-expert networks for data representation, it is necessary to calculate the trust level of a data point to different expert networks through a routing network, analogous to routing several copies of the data to different expert networks for calculation. This embodiment designs a multimodal routing network for this purpose. For different modalities of perceived data, dedicated routing networks are constructed for each modality (e.g., log routing network, indicator routing network, call link routing network). Simultaneously, a shared routing network is designed for extracting common semantics between modalities. When processing a data point, the multimodal routing network concatenates the shared routing network with the corresponding modality routing network based on the data modality. The two networks calculate the trust level of the data to different experts, and the results are merged to obtain the trust weights assigned to different expert networks. This embodiment trains a multi-path expert network to adapt to fault modes of various modalities of data. Combined with the multimodal routing network, the data is vectorized. The vector representations of different modalities of data are aligned in dimension and semantic features, allowing for aggregation. This enables the joint diagnosis of faults in the current time period by integrating the perceived data from various modalities. The multi-expert network representation based on multimodal routing proposed in this embodiment can vectorize and represent perceptual data of different modalities in a unified process. Through feature obfuscation and classification, it achieves joint multimodal fault diagnosis, rather than using independent fault diagnosis for each modality. This effectively avoids the ambiguity and inconsistency in fault diagnosis caused by separate representations of different modalities. The multi-expert network representation based on multimodal routing uses multiple expert networks, allowing different expert networks to have their own strengths and adapt to different modal data patterns. By weighted aggregation of multiple expert outputs through the multimodal routing network, it avoids the bias problem that when using a single expert to extract features from multiple modalities, the feature extraction of one modality is better, while the representation of other modalities is poor.

[0066] Furthermore, this embodiment also provides a computer system fault diagnosis system for multimodal sensing data, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute the computer system fault diagnosis method for the multimodal sensing data.

[0067] Those skilled in the art will understand that the technical solutions provided by this invention may take the form of a method, system, or computer program product. Therefore, this invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this invention may take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce an implementation of the flowchart... Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0068] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of a computer system using multimodal sensing data, characterized in that, Includes the following steps: S101, unify the vectorization representation of the multimodal sensing data of the computer system, including: for each modality of the multimodal sensing data, generate vectorized features using expert networks of multiple channels respectively; cascade the shared routing network and the modal routing network of the corresponding modality according to the input data modality to calculate the trust degree of the sensing data of the data modality to different expert networks; and obtain the feature representation of the data modality by weighted summing of the feature outputs of each expert network based on the trust degree. S102, feature obfuscation is performed on the feature representations of each modality to achieve superposition and unified feature fusion for diagnosis; S103, The multimodal fusion features obtained after feature confusion are fed into the classification network to obtain the fault diagnosis results of the computer system; In step S101, when the shared routing network and the corresponding modal routing network are concatenated according to the input data modality to calculate the trust level of a modality's data to different expert networks, the corresponding modal routing network includes a log routing network. Indicator routing network The call chain is determined by the network. And log routing network Indicator routing network The call chain is determined by the network. Share a global cross-modal routing network for use as a shared routing network The global cross-modal routing network The global cross-modal routing network is switched on. The modality routing network corresponding to the input data modality is connected so that only the corresponding modality routing network is activated when aggregating the outputs of multiple expert networks for a given data modality. The expression for calculating the trust level of a data modality with different expert networks is as follows: ； ； ； in, , and These represent the trust levels of different expert networks based on log modal data, indicator modal data, and call modal data, respectively. , and These represent the log modal data, indicator modal data, and call modal data sensed during the j-th system monitoring time interval, respectively. These are the weighting coefficients for shared routes and modal routes. For a global cross-modal routing network, Routing network for logs, For the indicator routing network, The call link is provided by the network.

2. The computer system fault diagnosis method based on multimodal sensing data according to claim 1, characterized in that, In step S101, the multimodal perception data of the computer system includes three modalities: log modal data, indicator modal data, and call modal data. The log modal data is the computer system's log data, which includes metadata server operation event records and object allocation and event records from the object storage server. The indicator modal data is the computer system's indicator data, which includes server temperature and voltage, and hardware resource utilization in the operating system, including CPU utilization, memory utilization, disk utilization, and network utilization. The call modal data is the call record data between microservices in the computer system. The fault diagnosis results include some or all of the following: network failure, process crash, login failure, access denied, and file loss.

3. The computer system fault diagnosis method based on multimodal sensing data according to claim 2, characterized in that, In step S101, when generating vectorized features using expert networks with multiple channels, the functional expression of the vectorized features generated by the expert networks with multiple channels is as follows: ； ； ； in, , and These are vectorized features generated by expert networks across multiple channels based on log modal data, indicator modal data, and call modal data. , and The first The system monitors log modal data, indicator modal data, and call modal data within a given time interval. ~ They are respectively number 1 to A network of experts across multiple channels. This represents the total number of expert networks.

4. The computer system fault diagnosis method based on multimodal sensing data according to claim 1, characterized in that, In step S101, the function expression for obtaining the feature representation of the data modality by weighted summation of the feature outputs of each expert network based on trust level is as follows: ； ； ； in, , and These are feature representations for log modal data, metric modal data, and call modal data, respectively. , and These represent the log modal data, indicator modal data, and call modal data sensed during the j-th system monitoring time interval, respectively. , and These represent the trust levels of different expert networks based on log modal data, indicator modal data, and call modal data, respectively. Represents element-wise product. , and These are vectorized features generated by expert networks across multiple channels based on log modal data, indicator modal data, and call modal data.

5. The computer system fault diagnosis method based on multimodal sensing data according to claim 1, characterized in that, In step S102, when the feature representations of each modality are obfuscated to achieve superposition, the method of firstly calculating the occurrence frequency of a modality's feature representation at the current time and the occurrence frequency in historical sensing data is used, and then calculating the weight of each modality's feature representation based on the occurrence frequency at the current time and the occurrence frequency in historical sensing data. Then, the feature representations of each modality are weighted and summed according to their weights to achieve the superposition of the feature representations of each modality.

6. The computer system fault diagnosis method based on multimodal sensing data according to claim 1, characterized in that, In step S103, when the features obtained after feature confusion are fed into the classification network to obtain the fault diagnosis results of the computer system, the classification network used is a three-layer perceptron.

7. A computer system fault diagnosis system for multimodal sensing data, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to perform a computer system fault diagnosis method for multimodal sensing data as described in any one of claims 1 to 6.

Images