Dpu failure prediction method and system

Abstract

The application discloses a DPU fault prediction method and system, relates to the technical field of computer fault prediction, and discloses the DPU fault prediction method and system, comprehensively collects in-band and out-of-band data, intelligently fuses and cross-verify, effectively improves prediction accuracy, can improve fault prediction accuracy, thereby reduces the false positive rate, realizes early fault identification and active intervention.

Description

Technical Field

[0001] This application relates to the field of computer fault prediction technology, and in particular to a method and system for predicting DPU faults. Background Technology

[0002] In data center and cloud computing infrastructure, the data processing unit (DPU) is a core computing component, and its operational stability directly impacts overall service continuity. With increasing DPU integration and functional complexity, the risks of hardware failures and business anomalies have significantly increased. Traditional operation and maintenance models rely on manual periodic inspections and passive fault responses, which inherently suffer from large monitoring blind spots and strong response delays, making early fault identification and proactive intervention difficult. Existing technical solutions attempt to predict faults through data collection, but generally have key limitations: some solutions only collect network traffic or storage access data within the business channel, ignoring physical layer information from hardware monitoring units and environmental sensors, resulting in a single dimension for fault cause analysis; other solutions, while simultaneously acquiring multi-source data, lack effective spatiotemporal alignment mechanisms and feature fusion strategies, leaving in-band business data and out-of-band basic data in a fragmented state, unable to uncover cross-dimensional data correlations. Especially in the data fusion stage, existing methods use fixed weight superposition or simple splicing, failing to implement adaptive weighting based on the dynamic importance of data, resulting in fused features that cannot accurately represent the true operating status of the DPU, ultimately leading to low fault prediction accuracy and a high false alarm rate.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main purpose of this application is to provide a DPU fault prediction method and system, which aims to improve the accuracy of fault prediction.

[0005] To achieve the above objectives, this application proposes a DPU fault prediction method, which includes:

[0006] The in-band business data set is obtained through the in-band data acquisition module, and the out-of-band basic data set is obtained through the out-of-band data acquisition module.

[0007] The in-band business data set and the out-of-band basic data set are input into the collaborative data preprocessing module for time alignment and intelligent fusion processing to obtain a collaborative feature data set.

[0008] The collaborative feature data set is input into the out-of-band model training module for model construction and training to obtain a fault prediction model.

[0009] The collaborative feature data set acquired in real time is input into the fault prediction model for fault prediction processing to obtain fault prediction results;

[0010] The fault prediction result, along with the real-time acquired in-band service data set and the real-time acquired out-of-band basic data set, are input into the collaborative fault detection module for bidirectional cross-validation to obtain the fault confirmation result.

[0011] Perform the corresponding fault handling operation based on the fault confirmation result.

[0012] In one embodiment, the steps of acquiring an in-band service data set through an in-band data acquisition module and acquiring an out-of-band basic data set through an out-of-band data acquisition module include:

[0013] The in-band data acquisition module collects network traffic data, stores access data, and encrypts and processes data in real time from the DPU's service processing unit to obtain the in-band service data set.

[0014] The out-of-band data acquisition module collects temperature, voltage, current, and frequency data in real time from the hardware monitoring unit and environmental sensors of the DPU to obtain the out-of-band basic data set.

[0015] In one embodiment, the step of inputting the in-band service data set and the out-of-band basic data set into a collaborative data preprocessing module for time alignment and intelligent fusion processing to obtain a collaborative feature data set includes:

[0016] The in-band service data set and the out-of-band basic data set are timestamped by the time alignment unit in the collaborative data preprocessing module to obtain the time-aligned in-band service data set and the time-aligned out-of-band basic data set.

[0017] The collaborative data preprocessing module uses an intelligent fusion unit to perform attention-based weighted fusion processing on the time-aligned in-band business data set and the time-aligned out-of-band basic data set to obtain the collaborative feature data set.

[0018] In one embodiment, the step of obtaining the collaborative feature data set by performing attention-based weighted fusion processing on the time-aligned in-band service data set and the time-aligned out-of-band basic data set through the intelligent fusion unit in the collaborative data preprocessing module includes:

[0019] The importance analysis subunit in the intelligent fusion unit calculates a first importance score for each data item in the time-aligned in-band service data set and a second importance score for each data item in the time-aligned out-of-band basic data set to obtain a set of first importance scores and a set of second importance scores.

[0020] The attention weighting subunit in the intelligent fusion unit performs differentiated weighting calculations on the time-aligned in-band business data set and the time-aligned out-of-band basic data set based on the first importance score set and the second importance score set to obtain a weighted fusion data set;

[0021] The feature extraction subunit in the intelligent fusion unit extracts multidimensional time series features from the weighted fusion data set to obtain the collaborative feature data set.

[0022] In one embodiment, the step of inputting the collaborative feature data set into the out-of-band model training module for model construction and training to obtain a fault prediction model includes:

[0023] The historical data analysis unit in the out-of-band model training module analyzes the fault modes in the collaborative feature dataset to determine the fault type classification and fault level division.

[0024] A time-series neural network model structure for fault prediction is constructed using the model building unit in the out-of-band model training module.

[0025] The time-series neural network model structure is trained by the model training unit in the out-of-band model training module using the collaborative feature data set labeled with fault information to obtain the fault prediction model.

[0026] In one embodiment, the method further includes:

[0027] The fault prediction model is incrementally trained and its parameters are optimized by the incremental learning unit in the out-of-band model training module based on the newly obtained collaborative feature data set, so as to obtain an optimized fault prediction model.

[0028] In one embodiment, the step of inputting the fault prediction result, the in-band service data set acquired in real time, and the out-of-band basic data set acquired in real time into the collaborative fault detection module for bidirectional cross-validation to obtain the fault confirmation result includes:

[0029] The in-band verification unit in the collaborative fault detection module compares the fault prediction result with the abnormal indicators in the real-time acquired in-band business data set to obtain the first-time consistency score.

[0030] The out-of-band verification unit in the collaborative fault detection module compares the fault prediction result with the abnormal indicators in the real-time acquired out-of-band basic data set to obtain a second time consistency score.

[0031] The comprehensive judgment unit in the collaborative fault detection module judges the reliability of the fault prediction result based on the first time consistency score and the second time consistency score, so as to obtain the fault confirmation result.

[0032] In one embodiment, the step of performing corresponding fault handling operations based on the fault confirmation result includes:

[0033] The fault level is determined by the fault level judgment unit in the fault response module based on the severity of the fault in the fault confirmation result.

[0034] The response strategy matching unit in the fault response module matches the corresponding automated processing strategy according to the fault level to obtain the automated processing strategy.

[0035] The execution unit in the fault response module performs the corresponding fault handling operation according to the automated processing strategy.

[0036] In one embodiment, the step of executing the corresponding fault handling operation according to the automated processing strategy by the execution unit in the fault response module includes:

[0037] When the fault level is the first level, the execution unit records the fault log and issues an early warning notification.

[0038] When the fault level is the second level, the execution unit records the fault log, issues an early warning notification, and automatically adjusts the system parameters.

[0039] When the fault level is level three, the execution unit records the fault log, issues an emergency notification, performs automatic system parameter adjustment, and initiates the backup DPU switching process.

[0040] Furthermore, to achieve the above objectives, this application also proposes a DPU fault prediction system, which includes: a memory, a processor, and a DPU fault prediction program stored in the memory and executable on the processor, wherein the DPU fault prediction program is configured to implement the steps of the DPU fault prediction method.

[0041] The DPU fault prediction method and system proposed in this application effectively improves prediction accuracy by comprehensively collecting in-band and out-of-band data and performing intelligent fusion and cross-validation. This improves the fault prediction accuracy, thereby reducing the false alarm rate and enabling early fault identification and proactive intervention. Attached Figure Description

[0042] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A flowchart illustrating an embodiment of the DPU fault prediction method of this application;

[0045] Figure 2 This is a schematic diagram of a structure provided for an embodiment of the DPU fault prediction system of this application.

[0046] Explanation of icon numbers:

[0047] 10. Memory; 20. Processor.

[0048] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0049] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of this application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0050] It should be understood that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0051] In existing technologies, DPU fault detection and handling methods are inefficient and struggle to accurately predict faults before they occur, leading to prolonged system downtime and increased risk of data loss. Existing methods suffer from incomplete data collection, unintelligent processing, and inaccurate predictions. For example, they rely on single-dimensional predictions or lack intelligent weight allocation, resulting in indistinct data features and an inability to accurately reflect the DPU's operational status.

[0052] Based on this, embodiments of this application provide a DPU fault prediction method, referring to... Figure 1 The DPU fault prediction method includes steps S100 to S600, wherein:

[0053] Step S100: Obtain the in-band business data set through the in-band data acquisition module, and obtain the out-of-band basic data set through the out-of-band data acquisition module;

[0054] Step S200: Input the in-band service data set and the out-of-band basic data set into the collaborative data preprocessing module for time alignment and intelligent fusion processing to obtain a collaborative feature data set;

[0055] Step S300: Input the collaborative feature data set into the out-of-band model training module for model construction and training to obtain a fault prediction model;

[0056] Step S400: Input the real-time acquired collaborative feature data set into the fault prediction model for fault prediction processing to obtain fault prediction results;

[0057] Step S500: Input the fault prediction result, the in-band service data set acquired in real time, and the out-of-band basic data set acquired in real time into the collaborative fault detection module for bidirectional cross-validation processing to obtain the fault confirmation result;

[0058] Step S600: Perform the corresponding fault handling operation based on the fault confirmation result.

[0059] In this embodiment, the DPU (Data Processing Unit) is a core component in data centers and cloud computing environments, primarily responsible for high-speed data processing, forwarding, and storage. The stable operation of the DPU is fundamental to the overall system performance and reliability. The in-band data acquisition module directly acquires data closely related to business logic from the DPU's business processing path. This data typically reflects the DPU's performance and status when executing specific business tasks. The out-of-band data acquisition module acquires data related to hardware operating status and environmental conditions from the DPU's non-business processing path. This data typically includes physical parameters such as temperature, voltage, and current, used to monitor the DPU's health. The in-band business data set refers to a series of data collected by the in-band data acquisition module that reflects the DPU's business processing activities. This set provides insights into the DPU's business layer. The out-of-band basic data set refers to a series of data collected by the out-of-band data acquisition module that reflects the DPU's hardware and environmental operating status. This set provides insights into the DPU's physical layer.

[0060] In this embodiment, the collaborative data preprocessing module is configured to unify the processing of data from different sources and formats, including time alignment and intelligent fusion, to eliminate data heterogeneity and enhance the correlation between data. The collaborative feature data set refers to the comprehensive feature data that integrates in-band business data and out-of-band basic data after processing by the collaborative data preprocessing module. This set aims to more comprehensively and accurately characterize the operating status of the DPU. The fault prediction model refers to the mathematical model trained by the out-of-band model training module, used to identify potential fault modes of the DPU and predict the probability of fault occurrence. The fault prediction result refers to the preliminary judgment on the possible faults of the DPU output by the fault prediction model after analyzing the real-time collaborative feature data. The collaborative fault detection module is configured to perform multi-dimensional verification of the fault prediction result, combining real-time in-band business data and out-of-band basic data to improve the accuracy and reliability of fault judgment. The fault confirmation result refers to the final determined DPU fault state and severity after bidirectional cross-validation by the collaborative fault detection module. Fault handling operations refer to a series of countermeasures taken based on the fault confirmation results, aimed at mitigating the impact of the fault, restoring system functions, or preventing the fault from worsening.

[0061] In this embodiment, the DPU fault prediction method first acquires an in-band service data set through an in-band data acquisition module, and then acquires an out-of-band basic data set through an out-of-band data acquisition module. Specifically, the in-band data acquisition module can be a software agent deployed in the DPU's operating system or virtualization layer. It periodically collects service-related performance metrics, such as the number of bytes sent / received by the network interface, CPU utilization, and memory usage, by intercepting or monitoring system calls and network interface activities. The out-of-band data acquisition module can be an independent hardware monitoring unit that communicates with the DPU's hardware sensors via a physical connection or dedicated management interface to obtain hardware-level operational data, such as periodically polling the temperature sensor on the DPU motherboard and the voltage / current readings of the power management unit.

[0062] Furthermore, the in-band business data set and the out-of-band basic data set are input into the collaborative data preprocessing module for time alignment and intelligent fusion processing to obtain a collaborative feature data set. Time alignment can be performed using timestamp-based matching, associating data points from different data sources according to their most recent timestamps. For example, for in-band and out-of-band data, data points occurring within a preset time window can be identified and treated as synchronous data. If no exact match is found, interpolation or nearest neighbor padding can be used for approximate alignment. Intelligent fusion processing can employ various data fusion techniques, such as simple feature concatenation, statistical methods, or machine learning methods. For example, the aligned in-band business data and out-of-band basic data can be directly concatenated into a longer feature vector, or common features can be extracted using dimensionality reduction techniques such as principal component analysis.

[0063] Based on this, the collaborative feature dataset is input into the out-of-band model training module for model building and training to obtain a fault prediction model. Model building may involve selecting appropriate machine learning algorithms, such as support vector machines, decision trees, random forests, or multilayer perceptrons. The model structure can be designed according to the data characteristics and prediction task requirements, such as determining the number of layers and neurons. Model training can employ supervised learning, utilizing historical collaborative feature datasets and their corresponding fault labels to optimize the model's internal parameters. During training, optimization algorithms such as gradient descent can be used, and model performance can be evaluated through techniques such as cross-validation.

[0064] Subsequently, the real-time acquired collaborative feature data set is input into the fault prediction model for fault prediction processing to obtain a fault prediction result. The real-time acquired collaborative feature data set refers to the feature data continuously generated and updated during the actual operation of the DPU through the aforementioned in-band and out-of-band data acquisition and collaborative data preprocessing procedures. Fault prediction processing refers to using this real-time feature data as input to the trained fault prediction model for inference calculations. Based on the patterns it has learned, the model outputs a predicted value about the current DPU state, such as a numerical value representing the probability of a fault, or a classification label indicating a specific fault type. The fault prediction result can be the raw predicted value output by the model, or a preliminary fault indication obtained after threshold judgment or post-processing.

[0065] Furthermore, the fault prediction result is input into the collaborative fault detection module for bidirectional cross-validation along with the real-time acquired in-band service data set and the real-time acquired out-of-band basic data set to obtain a fault confirmation result. The collaborative fault detection module can receive the fault prediction result and compare it with the real-time acquired raw in-band service data and out-of-band basic data. The bidirectional cross-validation process can include verification across multiple dimensions. For example, it can check whether the fault prediction result matches the abnormal behavior observed in the in-band service data. Simultaneously, it can also check whether the prediction result is correlated with hardware anomalies reflected in the out-of-band basic data. This mutual verification of multi-source information can improve the accuracy of fault diagnosis.

[0066] Finally, based on the fault confirmation result, corresponding fault handling operations are performed. The fault confirmation result may include information such as the type of fault, its severity, and the potential scope of impact. Fault handling operations can trigger preset automated scripts or manual intervention processes based on the severity and type of the fault confirmation result. For example, for minor faults, simply logging and issuing warnings may be sufficient; for moderate faults, attempts may be made to automatically adjust system parameters or restart some services; for severe faults, it may be necessary to start a backup system or perform an emergency shutdown.

[0067] In this embodiment, comprehensive DPU operational status characteristics are constructed by collaboratively collecting in-band service data and out-of-band basic data and performing intelligent fusion preprocessing. The fault prediction model trained based on these characteristics, combined with a two-way cross-validation mechanism, can effectively overcome the limitations of traditional methods such as incomplete data collection, unintelligent processing, and inaccurate prediction. This enables early and accurate prediction and confirmation of potential DPU faults, allowing for timely execution of corresponding fault handling operations, reducing system downtime risks, and ensuring the stable operation of the DPU.

[0068] In one feasible implementation, the steps of acquiring an in-band service data set through an in-band data acquisition module and acquiring an out-of-band basic data set through an out-of-band data acquisition module include: acquiring network traffic data, storage access data, and encrypted processing data in real time from the DPU's service processing unit through the in-band data acquisition module to acquire the in-band service data set; and acquiring temperature data, voltage data, current data, and frequency data in real time from the DPU's hardware monitoring unit and environmental sensors through the out-of-band data acquisition module to acquire the out-of-band basic data set.

[0069] In this embodiment, the in-band data acquisition module is responsible for obtaining data directly related to the actual workload from the core business processing unit of the DPU. The business processing unit of the DPU typically includes network interfaces, packet processors, storage controllers, and encryption / decryption engines, which are the foundation for the DPU to perform its core functions. Real-time acquired network traffic data can include packet throughput, latency, packet loss rate, connection status, etc., which directly reflect the DPU's performance at the network layer and potential congestion or anomalies. Storage access data covers the frequency of read / write operations to internal or external storage, I / O latency, cache hit rate, etc., which helps assess the health status and performance bottlenecks of the storage subsystem. Encryption processing data involves the rate of encryption / decryption operations, error rate, key management status, etc., which are crucial for ensuring data security and identifying security module failures. The in-band data acquisition module can be implemented as a software agent running on the DPU control plane or data plane, or as a performance monitoring unit integrated within the DPU hardware, acquiring the aforementioned real-time data by reading specific registers or counters.

[0070] In this embodiment, the out-of-band data acquisition module focuses on acquiring the physical operating status and environmental parameters of the DPU. The DPU's hardware monitoring unit typically integrates temperature sensors, voltage sensors, current sensors, and frequency monitoring circuitry, providing key physical parameters such as temperature, core voltage, supply current, and clock frequency for various regions within the chip. Environmental sensors monitor external conditions such as temperature and humidity of the chassis or environment where the DPU is located. Real-time temperature data reflects the DPU's heat dissipation and the presence of overheating risks; voltage and current data are directly related to the DPU's power consumption and power stability; abnormal fluctuations may indicate power failure or component damage; frequency data indicates the stability of the DPU's clock system; frequency drift or anomalies may lead to performance degradation or even system crashes. The out-of-band data acquisition module can be an independent microcontroller or dedicated monitoring chip, communicating with the DPU's hardware monitoring unit and external environmental sensors via standard interfaces (such as I2C or SMBus) to periodically read and record these physical parameters.

[0071] In this embodiment, through the above technical solution, this application clarifies the scope and source of data collection required for DPU fault prediction, ensuring the comprehensiveness and relevance of the acquired data. The collection of in-band service data sets enables the fault prediction model to directly perceive performance and behavioral anomalies of the DPU when executing core business functions, thereby effectively identifying faults related to software logic, network processing, or security functions. The collection of out-of-band basic data sets provides the model with direct evidence of the DPU's hardware health status and environmental impact, helping to discover potential faults caused by physical factors (such as overheating and power instability). This collaborative collection of in-band and out-of-band data not only compensates for the limitations of a single data source but also provides high-quality, multi-dimensional data support for subsequent collaborative data preprocessing and model training. Therefore, this solution significantly improves the accuracy and robustness of the fault prediction model, reduces the false alarm rate and false negative rate of fault prediction, and allows for more accurate assessment of the DPU's operating status, thus providing a solid data foundation for timely fault handling measures.

[0072] In one feasible implementation, the step of inputting the in-band service data set and the out-of-band basic data set into a collaborative data preprocessing module for time alignment and intelligent fusion processing to obtain a collaborative feature data set includes: performing timestamp alignment processing on the in-band service data set and the out-of-band basic data set through the time alignment unit in the collaborative data preprocessing module to obtain a time-aligned in-band service data set and a time-aligned out-of-band basic data set; and performing attention-based weighted fusion processing on the time-aligned in-band service data set and the time-aligned out-of-band basic data set through the intelligent fusion unit in the collaborative data preprocessing module to obtain the collaborative feature data set.

[0073] In this embodiment, the collaborative data preprocessing module aims to preprocess the original in-band business data set and out-of-band basic data set to eliminate data heterogeneity and provide high-quality, consistent input for subsequent fault prediction models. Its function is to standardize and integrate data from different data sources with different formats and temporal characteristics, ensuring data consistency in time and semantics. This module can be implemented by a software service running on the DPU control plane or as a dedicated processing unit within the DPU.

[0074] In this embodiment, the main function of the time alignment unit is to perform timestamp alignment processing on the in-band service data set and the out-of-band basic data set. Since in-band service data (such as network traffic) and out-of-band basic data (such as temperature and voltage) may be collected by different sensors or acquisition mechanisms at different frequencies and time points, precise time synchronization is required. Time alignment processing ensures that different types of data can be accurately correlated at the same point in time, thereby avoiding data misjudgment due to time misalignment. For example, interpolation algorithms (such as linear interpolation and spline interpolation) can be used to upsample low-sampling-rate data to high-sampling-rate data, or resampling techniques can be used to unify all data to a preset time granularity, ensuring the data integrity and consistency of each time step.

[0075] In this embodiment, the intelligent fusion unit is responsible for deeply fusing the time-aligned in-band service data set and the out-of-band basic data set to generate a more representative and predictive collaborative feature data set. Its "intelligence" lies in its ability to identify and utilize the inherent correlations between different data sources and to differentiate them based on their importance for fault prediction. The intelligent fusion unit can employ various fusion strategies. For example, in addition to the attention-based weighted fusion processing mentioned in subsequent embodiments, it can also use a deep learning-based multimodal fusion network to automatically extract high-level fusion features by learning the complex nonlinear relationships between different data modalities; or it can use statistical methods based on Bayesian networks or Kalman filtering to probabilistically fuse multi-source data to obtain a more robust joint state estimate.

[0076] In this embodiment, by introducing a collaborative data preprocessing module, this application can effectively solve the problem of inconsistency in the time dimension between the in-band business data set and the out-of-band basic data set. Specifically, the time alignment unit can accurately align the timestamps of data from different sources and with different sampling frequencies, ensuring that all data remain synchronized on the time axis, thereby avoiding analysis bias caused by data misalignment. Based on this, the intelligent fusion unit further employs an attention mechanism for weighted fusion processing, dynamically assigning different weights according to the contribution of different data features to fault prediction, thereby highlighting key information and suppressing noise and redundant information. This intelligent fusion method makes the generated collaborative feature data set not only highly consistent in time but also more refined and effective in information content, providing high-quality, highly relevant input for subsequent fault prediction models, significantly improving the accuracy and reliability of DPU fault prediction.

[0077] In one feasible implementation, the step of obtaining the collaborative feature data set by performing attention-based weighted fusion processing on the time-aligned in-band business data set and the time-aligned out-of-band basic data set through the intelligent fusion unit in the collaborative data preprocessing module includes: calculating a first importance score for each data item in the time-aligned in-band business data set and a second importance score for each data item in the time-aligned out-of-band basic data set through the importance analysis subunit in the intelligent fusion unit, to obtain a first importance score set and a second importance score set; performing differential weighting calculation on the time-aligned in-band business data set and the time-aligned out-of-band basic data set based on the first importance score set and the second importance score set through the attention weighting subunit in the intelligent fusion unit, to obtain a weighted fused data set; and extracting multi-dimensional time series features from the weighted fused data set through the feature extraction subunit in the intelligent fusion unit, to obtain the collaborative feature data set.

[0078] In this embodiment, the time-aligned in-band service data set and the time-aligned out-of-band basic data set are weighted and fused using an attention mechanism to obtain a collaborative feature data set. Specifically, the importance analysis subunit in the intelligent fusion unit calculates a first importance score for each data item in the time-aligned in-band service data set and a second importance score for each data item in the time-aligned out-of-band basic data set, thereby obtaining a set of first importance scores and a set of second importance scores. The importance analysis subunit aims to evaluate the relative contribution or correlation of each data item in the in-band service data and out-of-band basic data to DPU failure prediction. This can be achieved in various ways. For example, statistical methods, such as correlation analysis or mutual information, can be used to quantify the statistical dependence between each data item and known failure modes; or machine learning models, such as decision trees, random forests, or gradient boosting trees, can be used to calculate the importance score of each data item through feature importance evaluation algorithms (such as Gini impurity reduction, permutation importance, etc.). In addition, domain expert knowledge can be combined to assign higher initial importance scores to certain data items that are known to be closely related to DPU failures. The calculated first and second importance score sets can be normalized values ​​used to represent the potential influence of each data item in fault prediction.

[0079] Building upon this foundation, the attention-weighting subunit within the intelligent fusion unit performs differentiated weighting calculations on the time-aligned in-band business data set and the time-aligned out-of-band basic data set based on the first and second importance score sets, thereby obtaining a weighted fused data set. The role of the attention-weighting subunit is to assign different weights to different data items based on the results of importance analysis, thus highlighting key information and suppressing secondary or noisy information during data fusion. For example, a weighting function can be designed to map importance scores to weight coefficients, and then the original data items are multiplied by their corresponding weight coefficients to achieve weighting. Furthermore, self-attention or cross-attention mechanisms can be employed to dynamically learn the dependencies and importance between different data items and generate corresponding attention weights. For instance, an attention layer based on the Transformer architecture can be used to dynamically weight different data items through query, key, and value calculations, allowing data more critical to fault prediction to gain greater influence during the fusion process.

[0080] Subsequently, the feature extraction subunit within the intelligent fusion unit extracts multidimensional time-series features from the weighted fusion dataset to obtain a collaborative feature dataset. The feature extraction subunit aims to further refine more representative and discriminative features from the weighted data for use in subsequent fault prediction models. This can include various feature extraction techniques. For example, statistical features such as mean, variance, standard deviation, kurtosis, skewness, maximum, minimum, median, quartiles, energy, and entropy can be extracted to capture the basic statistical characteristics of the data. Time-domain features such as zero-crossing rate, peak factor, waveform factor, and impulse factor can also be extracted to reflect the data's variation over time. Furthermore, frequency-domain feature extraction methods, such as performing Fourier transforms on the data, can be used to extract the energy, dominant frequency, and bandwidth of the spectrum to analyze the frequency components of the data. More advanced feature extraction methods can employ deep learning models, such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs, such as long short-term memory networks (LSTM) or gated recurrent units (GRUs). These models can automatically learn and extract high-level, abstract time-series features from weighted fusion datasets, capturing complex patterns and temporal dependencies in the data. The resulting collaborative feature dataset is multidimensional and can comprehensively and precisely reflect the operational status of the DPU.

[0081] In this embodiment, by introducing an importance analysis subunit through the aforementioned technical solution, a refined importance assessment of various data points in both in-band business data and out-of-band basic data can be performed, quantifying their potential contribution to DPU failures. Based on this, an attention weighting subunit can differentiate the weighting of the data according to these importance scores, giving higher attention and weight to data more critical to failure prediction, effectively suppressing the influence of noise and irrelevant information. Subsequently, a feature extraction subunit extracts multi-dimensional time-series features from the weighted data, ensuring that the obtained collaborative feature dataset can more accurately and comprehensively reflect the actual operating status and potential failure modes of the DPU. This refined data fusion and feature extraction mechanism significantly improves the quality of the collaborative feature dataset, thereby providing more reliable input for the subsequent failure prediction model, improving the accuracy and sensitivity of DPU failure prediction, and reducing the risk of false positives and false negatives.

[0082] In one feasible implementation, the step of inputting the collaborative feature data set into an out-of-band model training module for model construction and training to obtain a fault prediction model includes: analyzing the fault patterns in the collaborative feature data set through the historical data analysis unit in the out-of-band model training module to determine fault type classification and fault level division; constructing a time-series neural network model structure for fault prediction through the model construction unit in the out-of-band model training module; and training the time-series neural network model structure using the collaborative feature data set labeled with fault information through the model training unit in the out-of-band model training module to obtain the fault prediction model.

[0083] In this embodiment, the historical data analysis unit in the out-of-band model training module is responsible for in-depth mining and analysis of the long-accumulated collaborative feature data set. Its role is to identify potential patterns, abnormal behavioral characteristics, and pre-failure signs related to DPU failures from massive amounts of historical data. Failure pattern analysis can be achieved through various data analysis techniques. For example, statistical analysis methods such as mean, variance, and correlation analysis can be used to discover abnormal fluctuations in data characteristics; clustering algorithms can be used for unsupervised learning of historical data to automatically discover different failure clusters or abnormal behavioral patterns; or expert experience and rule engines can be combined to define known failure characteristics and thresholds. Based on the analysis results of failure patterns, this unit further clarifies the types of failures that may occur in the DPU and classifies the severity of each failure. Failure type classification can use supervised learning methods, such as support vector machines, decision trees, or random forests, to train a classifier on labeled failure data. Failure level classification can be quantified based on factors such as the degree of impact of the failure on system performance, the ease of recovery, and potential harm, through threshold setting, expert evaluation, or regression models.

[0084] Based on this, the model building unit in the out-of-band model training module is responsible for designing and building a suitable machine learning model architecture according to the fault modes and prediction targets determined by the historical data analysis unit. Its core task is to select and configure a model that can effectively process time-series data, capture temporal dependencies, and complex nonlinear features. Given that collaborative feature datasets typically exhibit significant time-series characteristics, i.e., there is a temporal order and dependency between data points, this application adopts a time-series neural network model structure. Such models are particularly adept at processing sequential data and can learn and memorize long-term dependencies in the data. Specific model structures may include, but are not limited to: Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), One-Dimensional Convolutional Neural Networks (1D-CNNs), or models based on the Transformer architecture. For example, LSTM and GRU, through their unique gating mechanisms, can effectively alleviate the gradient vanishing or gradient exploding problems in traditional Recurrent Neural Networks (RNNs), thereby better capturing dependencies over long time spans. The model building unit selects and optimizes hyperparameters such as the number of layers, number of neurons, activation function, and regularization method of these neural networks based on the characteristics of the actual data and the complexity of the prediction task, in order to form an efficient and robust fault prediction model structure.

[0085] Subsequently, the model training unit in the out-of-band model training module is responsible for executing the training process of the machine learning model, that is, adjusting the model's internal parameters through iterative optimization algorithms so that it can learn the ability to predict targets from the input data. The collaborative feature dataset labeled with fault information is a key input for model training. This data, during the historical data analysis phase, is labeled with corresponding fault type and / or fault level labels for each time point or time period in the collaborative feature dataset through manual labeling, semi-supervised learning, or combining known fault event logs. These labels are the basis for the model's prediction learning, enabling the model to associate specific data patterns with specific fault states. The process of training the time series neural network model structure typically involves the following steps: First, the labeled collaborative feature dataset is divided into a training set, a validation set, and a test set. Then, the model training unit uses a backpropagation algorithm and an optimizer to minimize the error between the model's predicted output and the true fault labels. During training, the model's performance is monitored through the validation set to prevent overfitting, and training hyperparameters such as the learning rate and batch size are adjusted as needed. Training continues until the model's performance on the validation set reaches a preset standard or converges. After sufficient training and optimization, the model training unit will output a time-series neural network model with fixed parameters and predictive capabilities, namely the fault prediction model. This model can accept new collaborative feature datasets as input and output corresponding fault type classifications and fault level divisions, thereby realizing fault prediction by the DPU.

[0086] In this embodiment, when the collaborative feature dataset is input into the out-of-band model training module for model construction and training, the historical data analysis unit first performs an in-depth analysis of the fault patterns in the collaborative feature dataset, thereby accurately determining the fault type classification and fault level division. This preprocessing step provides clear prediction targets and label information for subsequent model construction, avoiding blind training and significantly improving the model's ability to identify different fault scenarios. Based on this, the model construction unit specifically constructs a time-series neural network model structure for fault prediction. Given the inherent temporal characteristics of DPU runtime data, using a time-series neural network can effectively capture dynamic changes and long-term dependencies in the data, thereby more accurately identifying fault precursors. Finally, the model training unit uses the collaborative feature dataset labeled with fault information to train the constructed time-series neural network model structure, ensuring that the model can learn reliable prediction patterns from real fault data. Through this systematic model building and training process, this application can obtain a high-precision and robust fault prediction model. This model can not only accurately predict the types of faults that may occur in the DPU, but also finely assess the severity of the faults, providing a solid foundation for subsequent fault prediction results and bidirectional cross-validation. This significantly improves the accuracy and reliability of DPU fault prediction and effectively supports the stable operation and timely maintenance of the DPU.

[0087] In one feasible implementation, the method further includes: performing incremental training and parameter optimization on the fault prediction model based on the subsequently obtained new collaborative feature data set through the incremental learning unit in the out-of-band model training module, so as to obtain an optimized fault prediction model.

[0088] In this embodiment, the main function of the incremental learning unit is to enable the fault prediction model to continuously learn from new data without having to retrain the entire model from scratch each time. This is crucial for handling continuously generated data streams, adapting to dynamic changes in the DPU operating environment, and ensuring the timeliness and accuracy of model predictions. This incremental learning unit can be implemented using various algorithms. For example, it can be based on online learning algorithms, using mini-batch gradient descent or its variants (such as the Adam optimizer) to periodically update the model parameters in small increments. In practical applications, the incremental learning unit can also include a data caching mechanism to temporarily store the collaborative feature data set output by the collaborative data preprocessing module within a recent period, so that incremental training can be triggered when preset conditions (such as data volume thresholds or time intervals) are met.

[0089] In this embodiment, the subsequently obtained new collaborative feature data set refers to the latest data collected in real time or periodically by the collaborative data preprocessing module during the continuous operation of the system after the fault prediction model has completed its initial training and been put into use. This data can promptly reflect the latest information on the current and recent operating status and environment of the DPU, serving as the basis for the model to self-adjust and adapt to new situations. These new collaborative feature data sets can be continuously generated by the collaborative data preprocessing module and received and processed by the incremental learning unit at a predetermined frequency (e.g., hourly, daily, or when the data volume accumulates to a certain level).

[0090] In this embodiment, the incremental training and parameter optimization process refers to the gradual adjustment and updating of the model based on the existing fault prediction model parameters using the aforementioned new collaborative feature data set, rather than a complete retraining. Parameter optimization involves iteratively adjusting the model's internal parameters, such as weights and biases, to better capture new data distribution characteristics and potential fault modes, thereby continuously improving the model's predictive performance. When implementing incremental training, the following strategies can be adopted: First, the new collaborative feature data set is divided into several mini-batch datasets. The model's loss function is calculated using one batch of data at a time, and the model parameters are updated based on the loss value. Second, to avoid excessive impact of new data on the existing model's knowledge, a smaller learning rate is typically used for parameter updates. Furthermore, regularization techniques (such as L1 or L2 regularization) can be combined to prevent overfitting on new data while maintaining its generalization ability to historical data. In some advanced implementations, knowledge distillation techniques can also be introduced, using the old model as the "teacher model" and the new model as the "student model." Through a specific distillation loss function, the new model can effectively retain and inherit the experiential knowledge of the old model while learning new knowledge.

[0091] In this embodiment, the optimized fault prediction model obtained after the above incremental training and parameter optimization processes can more accurately capture the latest operating status and potential fault modes of the DPU. This optimized model will replace or update the original fault prediction model for subsequent real-time fault prediction processing, thereby ensuring that the system always uses the latest and most accurate model for fault prediction.

[0092] In this embodiment, through the above technical solution, the fault prediction model can continuously adapt to the dynamic changes in the DPU operating environment and fault modes. The introduction of incremental learning units allows the model to update its parameters by continuously absorbing new collaborative feature data sets without frequent, time-consuming, and resource-intensive full retraining. This effectively solves the performance degradation problem caused by data distribution drift or the emergence of new fault modes, ensuring the accuracy and timeliness of fault prediction. Therefore, this application can significantly improve the robustness and reliability of the DPU fault prediction system, reduce the risk of false alarms or false negatives due to model obsolescence, and thus improve the overall operation and maintenance efficiency and stability of the DPU system.

[0093] In one feasible implementation, the step of inputting the fault prediction result, the in-band service data set acquired in real time, and the out-of-band basic data set acquired in real time into a collaborative fault detection module for bidirectional cross-validation to obtain a fault confirmation result includes: comparing the fault prediction result with the abnormal indicators in the in-band service data set acquired in real time through the in-band verification unit in the collaborative fault detection module to obtain a first time consistency score; comparing the fault prediction result with the abnormal indicators in the out-of-band basic data set acquired in real time through the out-of-band verification unit in the collaborative fault detection module to obtain a second time consistency score; and judging the reliability of the fault prediction result based on the first time consistency score and the second time consistency score through the comprehensive judgment unit in the collaborative fault detection module to obtain the fault confirmation result.

[0094] In this embodiment, the collaborative fault detection module plays a crucial verification role, aiming to verify the preliminary fault prediction results from multiple dimensions and sources, thereby improving the accuracy and reliability of fault judgment. Specifically, the in-band verification unit is responsible for comparing the fault prediction results with anomaly indicators in the real-time acquired in-band service data set. For example, if the fault prediction result indicates that a certain DPU may have a network processing anomaly, the in-band verification unit will check whether there are any service-level anomaly indicators in the real-time network traffic data, storage access data, or encryption processing data that match the anomaly, such as traffic spikes, increased access latency, or increased encryption failure rates. This comparison process can be performed by calculating correlation, deviation degree, or anomaly pattern matching degree, thereby quantifying and generating a first-time consistency score, which reflects the degree of consistency between the fault prediction result and the anomaly in the service data.

[0095] In this embodiment, the out-of-band verification unit compares the fault prediction result with anomalies in the real-time acquired out-of-band basic data set. For example, if the fault prediction result indicates that the DPU may have a hardware fault, the out-of-band verification unit checks whether there are any hardware-level anomalies related to the fault, such as excessively high temperature, unstable voltage, abnormal current, or frequency drift, in the real-time temperature data, voltage data, current data, or frequency data. Similar to in-band verification, this comparison process can also be performed using methods such as statistical analysis, threshold judgment, or pattern recognition to obtain a second time consistency score, which characterizes the degree of agreement between the fault prediction result and the hardware basic data anomalies.

[0096] In this embodiment, the comprehensive judgment unit comprehensively evaluates the reliability of the fault prediction result based on the first time consistency score and the second time consistency score. For example, the comprehensive judgment unit can use methods such as weighted average, logical judgment rules (such as "AND" and "OR" relationships), or machine learning classifiers to perform fusion analysis on the two consistency scores. If both scores reach a preset threshold, or one score is significantly higher than the threshold and the other score is not lower than the lower threshold, the fault prediction result can be considered to have high reliability, and a fault confirmation result can be output. Conversely, if both scores are low, it may indicate that the fault prediction result is a false alarm or uncertain, and a confirmation result of "no fault" or "further observation required" can be output.

[0097] In this embodiment, through the aforementioned bidirectional cross-validation process, this application effectively addresses the potential reliability issues of single-model prediction results. By introducing in-band business data and out-of-band basic data for mutual verification, even if noise or model prediction deviations occur in one type of data source, verification from another type of data source can correct or confirm the issue, significantly reducing the risk of false positives and false negatives. The acquisition of the first and second time-consistency scores provides a quantitative basis for comprehensive judgment, making the fault confirmation process more objective and accurate. This multi-source, multi-dimensional verification mechanism ensures the accuracy and robustness of fault confirmation results, providing a solid foundation for subsequent fault handling operations, avoiding resource waste or fault escalation due to misjudgment, thereby improving the overall stability and operational efficiency of the DPU system.

[0098] In one feasible implementation, the step of performing corresponding fault handling operations based on the fault confirmation result includes: determining the fault level by the fault level judgment unit in the fault response module based on the fault severity in the fault confirmation result; matching the corresponding automated handling strategy by the response strategy matching unit in the fault response module based on the fault level to obtain the automated handling strategy; and performing the corresponding fault handling operation by the execution unit in the fault response module according to the automated handling strategy.

[0099] In this embodiment, the fault response module is a software or hardware functional unit used to handle confirmed faults. This module receives fault confirmation results from the collaborative fault detection module and is responsible for coordinating subsequent fault handling processes to ensure that the DPU system can respond to faults in a timely and appropriate manner. The fault level judgment unit is a component of the fault response module. Its main function is to parse the fault severity information contained in the fault confirmation results and map it to predefined fault levels. For example, this unit can classify faults into different levels, such as "general," "important," or "urgent," based on multiple dimensions such as the scope of the fault's impact on business, its duration, and the difficulty of recovery. This can be achieved through a built-in rule engine, an expert knowledge-based decision tree, or a pre-trained classification model to ensure the accuracy and consistency of fault level determination. The response strategy matching unit is also part of the fault response module. Its responsibility is to select the most suitable automated processing strategy from a preset strategy library based on the fault level determined by the fault level judgment unit. This strategy library stores standardized processing procedures for different fault levels and types. For example, for minor faults, it may only be necessary to log and issue warnings, while for severe faults, it may be necessary to immediately isolate the affected DPU and activate the backup system. This unit can be implemented using lookup tables, rule-based matching algorithms, or more sophisticated machine learning recommendation systems to ensure that the matching strategy optimizes fault handling to the greatest extent possible. The execution unit is the final executor of the fault response module. It receives the automated processing strategy output by the response strategy matching unit and translates it into specific system operation instructions to perform fault handling operations on the DPU or related systems. This may include calling system APIs, sending control commands, triggering the execution of preset scripts, adjusting system parameters, and activating backup resources. This unit typically integrates various system management tools and interfaces, enabling interaction with the DPU's underlying hardware and upper-level software to ensure that the automated processing strategy is executed accurately.

[0100] In this embodiment, by introducing a fault response module and refining the fault handling process, this application enables the fault level judgment unit to accurately determine the fault level based on the severity of the fault in the fault confirmation results. Based on this, the response strategy matching unit can match the most appropriate automated processing strategy for different fault levels, avoiding a one-size-fits-all approach to all faults. Finally, the execution unit executes the corresponding fault handling operation according to the matched automated processing strategy, thereby achieving intelligent, automated, and refined fault response. This hierarchical response mechanism significantly improves the efficiency and accuracy of fault handling, effectively reduces the risk of misoperation, and ensures that the DPU system can recover quickly and effectively in the face of various faults, maximizing business continuity and system stability.

[0101] In one feasible implementation, the steps of executing corresponding fault handling operations by the execution unit in the fault response module according to the automated processing strategy include: when the fault level is the first level, the execution unit records the fault log and issues an early warning notification; when the fault level is the second level, the execution unit records the fault log, issues an early warning notification, and performs automatic system parameter adjustment processing; when the fault level is the third level, the execution unit records the fault log, issues an emergency notification, performs automatic system parameter adjustment processing, and initiates the backup DPU switching process.

[0102] In this embodiment, a Level 1 fault typically refers to a potential, minor anomaly or performance fluctuation that has not yet significantly affected the normal operation of the DPU. At this stage, fault logs are recorded by the execution unit to provide historical data support for subsequent fault analysis and system optimization. Issuing early warning notifications alerts maintenance personnel to potential risks, enabling timely manual investigation or preventative maintenance to prevent further escalation of the fault. Fault log recording is a fundamental step in fault handling, detailing the time, type, relevant parameters, system status, and countermeasures taken during the fault's occurrence. This log data is invaluable for fault tracing, performance bottleneck location, system optimization, and iterative training of future fault prediction models. The early warning notification mechanism aims to send alert information to pre-defined maintenance personnel or system administrators through various channels (such as SMS, email, instant messaging tools, and alarm platforms). Its purpose is to promptly inform relevant parties at the initial stage of a fault or potential risk phase, allowing them to intervene in a timely manner for manual inspection, confirmation, or preventative measures, thereby preventing fault escalation.

[0103] In this embodiment, a Level 2 fault indicates that the DPU system has experienced significant anomalies or performance degradation, potentially affecting some business functions or service quality. In this situation, in addition to recording fault logs and issuing warning notifications, the execution unit will also automatically adjust system parameters. For example, it can automatically adjust parameters such as DPU resource allocation, load balancing strategies, cache size, or network bandwidth to attempt to mitigate the impact of the fault, restore system performance, or buy time for manual intervention. Automatic system parameter adjustment is a proactive fault mitigation or self-healing mechanism. When a specific fault mode is detected, the system can automatically modify the DPU's operating parameters according to preset strategies or machine learning models. For example, it can reduce the priority of certain non-critical tasks, adjust power consumption modes, and optimize data paths to reduce system load, avoid fault points, or optimize resource utilization, thereby restoring or maintaining system performance without interrupting service.

[0104] In this embodiment, a Level 3 fault represents a severe failure or impending service interruption in the DPU system, posing a significant threat to business continuity. At this time, in addition to recording fault logs and automatically adjusting system parameters, the execution unit will issue an emergency notification with the highest priority to relevant personnel for emergency handling. More importantly, the system will initiate a backup DPU switchover process, meaning that business traffic on the currently failed DPU will be quickly and seamlessly switched to a pre-configured backup DPU to maximize business continuity and availability and avoid prolonged service interruptions. The emergency notification is an immediate alert for high-priority, high-severity faults. It typically triggers a higher-level response process, such as calling on-duty personnel or activating the emergency response team, ensuring human intervention is obtained in the shortest possible time to address serious issues that could lead to business interruption. Initiating the backup DPU switchover process is a key measure to ensure the high availability of the DPU system. When the primary DPU experiences a severe failure, the system will automatically or semi-automatically switch all or part of the business traffic from the failed DPU to the healthy backup DPU. This process typically includes steps such as fault detection, verification of backup DPU status, updating routing information, and data synchronization to ensure that services transition as smoothly as possible during the switchover process and minimize service interruption time.

[0105] In this embodiment, through the above technical solution, this application can classify and respond to DPU faults according to the severity of the fault reflected in the fault confirmation results. This classified processing mechanism avoids a one-size-fits-all approach to all faults, thereby achieving more refined and efficient fault management. For minor Level 1 faults, simply logging and issuing early warnings preserves fault information for subsequent analysis while avoiding unnecessary resource consumption and interference with system operation. For moderate Level 2 faults, automatic adjustment of system parameters is added on top of early warnings, which can promptly mitigate the impact of the fault, prevent the problem from escalating, and effectively improve the system's self-healing capability and stability. For severe Level 3 faults, emergency notifications and the initiation of a backup DPU switchover process ensure rapid service recovery at the most critical moment, maximizing the continuity and high availability of DPU services. This differentiated response strategy significantly improves the intelligence level of DPU fault prediction and handling, optimizes operational efficiency, and greatly reduces the risk and loss of business interruption caused by faults.

[0106] In the embodiments of this application, the DPU fault prediction method effectively improves prediction accuracy by comprehensively collecting in-band and out-of-band data and performing intelligent fusion and cross-validation, thereby improving the fault prediction accuracy rate, reducing the false alarm rate, and realizing early fault identification and proactive intervention.

[0107] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the DPU fault prediction method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0108] This application also provides a DPU fault prediction system, referenced Figure 2 The DPU fault prediction system includes: a memory 10, a processor 20, and a DPU fault prediction program stored on the memory 10 and executable on the processor 20. The DPU fault prediction program is configured to implement the steps of the DPU fault prediction method.

[0109] The DPU fault prediction system provided in this application, employing the DPU fault prediction method in the above embodiments, can improve the accuracy of fault prediction. Compared with the prior art, the beneficial effects of the DPU fault prediction system provided in this application are the same as those of the DPU fault prediction method provided in the above embodiments, and other technical features of the DPU fault prediction system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0110] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0111] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. All equivalent structural transformations made under the technical concept of this application using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of this application.

Claims

1. A method for predicting DPU faults, characterized in that, The DPU fault prediction method includes: The in-band business data set is obtained through the in-band data acquisition module, and the out-of-band basic data set is obtained through the out-of-band data acquisition module. The in-band business data set and the out-of-band basic data set are input into the collaborative data preprocessing module for time alignment and intelligent fusion processing to obtain a collaborative feature data set. The collaborative feature data set is input into the out-of-band model training module for model construction and training to obtain a fault prediction model. The collaborative feature data set acquired in real time is input into the fault prediction model for fault prediction processing to obtain fault prediction results; The fault prediction result, along with the real-time acquired in-band service data set and the real-time acquired out-of-band basic data set, are input into the collaborative fault detection module for bidirectional cross-validation to obtain the fault confirmation result. Perform the corresponding fault handling operation based on the fault confirmation result; The steps of inputting the in-band business data set and the out-of-band basic data set into the collaborative data preprocessing module for time alignment and intelligent fusion processing to obtain the collaborative feature data set include: The in-band service data set and the out-of-band basic data set are timestamped by the time alignment unit in the collaborative data preprocessing module to obtain the time-aligned in-band service data set and the time-aligned out-of-band basic data set. The collaborative data preprocessing module uses an intelligent fusion unit to perform attention-based weighted fusion processing on the time-aligned in-band business data set and the time-aligned out-of-band basic data set to obtain the collaborative feature data set. The steps of obtaining the collaborative feature data set by performing attention-based weighted fusion processing on the time-aligned in-band business data set and the time-aligned out-of-band basic data set through the intelligent fusion unit in the collaborative data preprocessing module include: The importance analysis subunit in the intelligent fusion unit calculates a first importance score for each data item in the time-aligned in-band service data set and a second importance score for each data item in the time-aligned out-of-band basic data set to obtain a set of first importance scores and a set of second importance scores. The attention weighting subunit in the intelligent fusion unit performs differentiated weighting calculations on the time-aligned in-band business data set and the time-aligned out-of-band basic data set based on the first importance score set and the second importance score set to obtain a weighted fusion data set; The feature extraction subunit in the intelligent fusion unit extracts multidimensional time series features from the weighted fusion data set to obtain the collaborative feature data set. The steps for inputting the fault prediction result, the in-band service data set acquired in real time, and the out-of-band basic data set acquired in real time into the collaborative fault detection module for bidirectional cross-validation to obtain the fault confirmation result include: The in-band verification unit in the collaborative fault detection module compares the fault prediction result with the abnormal indicators in the real-time acquired in-band business data set to obtain the first-time consistency score. The out-of-band verification unit in the collaborative fault detection module compares the fault prediction result with the abnormal indicators in the real-time acquired out-of-band basic data set to obtain a second time consistency score. The comprehensive judgment unit in the collaborative fault detection module judges the reliability of the fault prediction result based on the first time consistency score and the second time consistency score, so as to obtain the fault confirmation result.

2. The DPU fault prediction method as described in claim 1, characterized in that, The steps of acquiring in-band business data set through in-band data acquisition module and acquiring out-of-band basic data set through out-of-band data acquisition module include: The in-band data acquisition module collects network traffic data, stores access data, and encrypts and processes data in real time from the DPU's service processing unit to obtain the in-band service data set. The out-of-band data acquisition module collects temperature, voltage, current, and frequency data in real time from the hardware monitoring unit and environmental sensors of the DPU to obtain the out-of-band basic data set.

3. The DPU fault prediction method as described in claim 1, characterized in that, The steps of inputting the collaborative feature dataset into the out-of-band model training module for model construction and training to obtain a fault prediction model include: The historical data analysis unit in the out-of-band model training module analyzes the fault modes in the collaborative feature dataset to determine the fault type classification and fault level division. A time-series neural network model structure for fault prediction is constructed using the model building unit in the out-of-band model training module. The time-series neural network model structure is trained by the model training unit in the out-of-band model training module using the collaborative feature data set labeled with fault information to obtain the fault prediction model.

4. The DPU fault prediction method as described in claim 3, characterized in that, The method further includes: The fault prediction model is incrementally trained and its parameters are optimized by the incremental learning unit in the out-of-band model training module based on the newly obtained collaborative feature data set, so as to obtain an optimized fault prediction model.

5. The DPU fault prediction method as described in claim 1, characterized in that, The steps for performing corresponding fault handling operations based on the fault confirmation result include: The fault level is determined by the fault level judgment unit in the fault response module based on the severity of the fault in the fault confirmation result. The response strategy matching unit in the fault response module matches the corresponding automated processing strategy according to the fault level to obtain the automated processing strategy. The execution unit in the fault response module performs the corresponding fault handling operation according to the automated processing strategy.

6. The DPU fault prediction method as described in claim 5, characterized in that, The steps for the execution unit in the fault response module to perform the corresponding fault handling operation according to the automated processing strategy include: When the fault level is the first level, the execution unit records the fault log and issues an early warning notification. When the fault level is the second level, the execution unit records the fault log, issues an early warning notification, and automatically adjusts the system parameters. When the fault level is level three, the execution unit records the fault log, issues an emergency notification, performs automatic system parameter adjustment, and initiates the backup DPU switching process.

7. A DPU fault prediction system, characterized in that, The DPU fault prediction system includes: a memory, a processor, and a DPU fault prediction program stored in the memory and executable on the processor, the DPU fault prediction program being configured to implement the steps of the DPU fault prediction method as described in any one of claims 1 to 6.

Images