Operation and maintenance management method and system of data center

Abstract

The invention discloses an operation and maintenance management method and system of a data center, and relates to the field of operation and maintenance management of the data center. The method comprises the following steps: acquiring a Redfish protocol response message and a time sequence feature sequence of a multi-manufacturer substrate management controller; performing multi-modal fusion analysis on the Redfish protocol response message and the time sequence feature sequence by adopting a double-flow self-supervised contrast learning algorithm, and extracting a comprehensive semantic deviation feature vector of a manufacturer private extension protocol and a standard protocol; performing time sequence analysis on the time sequence feature sequence by using a time sequence attention mask mechanism, and identifying a time sequence sensitive interval in a protocol interaction process; constructing a dynamic protocol knowledge base according to the comprehensive semantic deviation feature vector and the time sequence sensitive interval; and according to the dynamic protocol knowledge base, generating an operation and maintenance strategy of the substrate management controller equipment to be upgraded. According to the method, the Redfish protocol semantic features and the time sequence behavior features are fused, the manufacturer-differentiated protocol knowledge base is constructed, and the self-adaptive firmware upgrading operation and maintenance strategy is generated.

Description

Technical Field

[0001] This application relates to the field of data center operation and maintenance management, and in particular to a data center operation and maintenance management method and system. Background Technology

[0002] With the continuous expansion of data center scale and the increasing complexity of business operations, modern large-scale data centers generally adopt a hybrid deployment architecture with multi-brand servers. This achieves a flexible balance between procurement costs, supply chain stability, and performance indicators by introducing hardware equipment from different vendors. However, this heterogeneous deployment strategy has caused serious compatibility issues in critical operation and maintenance scenarios such as firmware upgrades and equipment configuration changes. Although the Baseboard Management Controllers (BMCs) of various vendors nominally follow the Redfish standard protocol released by the DMTF organization, in actual implementations, they generally have a large number of proprietary extended fields and non-standardized protocol interaction sequences. This makes it difficult for unified automated operation and maintenance tools to adapt to the differentiated behavior patterns of multi-vendor equipment. Operation and maintenance personnel have to maintain independent upgrade scripts and policy rules for different brands of equipment, which significantly increases operation and maintenance costs and the risk of human error.

[0003] At the engineering implementation level of the Redfish protocol, mainstream vendors (such as HP iLO, Dell iDRAC, Inspur BMC, etc.) exhibit significant semantic deviations and timing differences in their extensions to the standard protocol. These differences have become the core technical bottleneck restricting the automation of firmware upgrades.

[0004] Various manufacturers have introduced a large number of proprietary extended fields located under the OEM node in the JSON structure of the Redfish protocol response message. For example, HP iLO has added a proprietary "HPE.FirmwareStatus" status code system to the Firmware Inventory resource, while Dell iDRAC uses the "Dell.Update Service.Job Queue" mechanism to manage upgrade tasks. The semantic definitions of these proprietary fields deviate from the DMTF standard specification, causing standardized parsing tools to be unable to accurately identify the actual operating status of the device.

[0005] Furthermore, different manufacturers' BMCs exhibit millisecond-level fine-grained differences in their control logic for protocol interaction timing during firmware upgrades. Specifically: HP iLO v2.3 requires a 300±50ms power policy negotiation window to be inserted before sending the Firmware Update command to complete power state pre-negotiation; Dell iDRAC v4.2 mandates continuous heartbeat packet transmission during the firmware transfer phase (with a transmission interval of ≤2 seconds), otherwise, a transmission interruption will be triggered; Inspur BMC's command response latency during the verification and confirmation phase ranges from 800-1200ms, significantly higher than the standard protocol's recommended 500ms threshold. These manufacturer-specific timing-sensitive windows and response latency differences lead to uncontrollable state shifts in the protocol state machines of different brands of devices when performing the same batch of upgrade operations using upgrade tools with uniform timeout configurations. Some devices falsely report failures due to premature timeout judgments, while others fail to complete the upgrade due to insufficient waiting time.

[0006] Currently, the industry mainly relies on manual experience to maintain the upgrade strategy configuration files of various manufacturers' equipment, or uses rule-based conditional judgment logic to handle manufacturer differences. However, these methods have the following drawbacks: (1) They cannot automatically identify the protocol behavior patterns of new manufacturers' equipment or new firmware versions, and adapting to new equipment requires a lot of manual testing and rule adjustments; (2) They lack the ability to quantitatively analyze protocol semantic deviations and timing differences, making it difficult to accurately locate the key time windows that affect the success rate of upgrades; (3) Static configuration rules cannot be dynamically updated with the iteration of device firmware, and maintenance costs increase exponentially with the increase of device type. Therefore, there is an urgent need for an intelligent operation and maintenance management method that can automatically learn the characteristics of manufacturer protocol differences, dynamically build a knowledge base, and generate adaptive operation and maintenance strategies. Summary of the Invention

[0007] To address the poor firmware upgrade compatibility caused by differences in protocols among multiple vendors' baseboard management controllers in existing technologies, this application provides a data center operation and maintenance management method and system. By integrating the semantic features and timing behavior features of the Redfish protocol, a vendor-differentiated protocol knowledge base is constructed to generate adaptive firmware upgrade operation and maintenance strategies.

[0008] One aspect of this application provides a data center operation and maintenance management method, comprising: S1, collecting Redfish protocol response messages and timing feature sequences from multi-vendor baseboard management controllers; S2, employing a dual-stream self-supervised contrastive learning algorithm to perform multimodal fusion analysis on the Redfish protocol response messages and timing feature sequences, extracting a comprehensive semantic deviation feature vector between vendor-specific extended protocols and standard protocols; wherein, the dual-stream self-supervised contrastive learning algorithm includes a protocol stream encoder and a timing stream encoder, the protocol stream encoder extracting protocol semantic features based on the BERT architecture, and the timing stream encoder extracting timing behavior features based on a temporal convolutional network; S3, utilizing a timing attention masking mechanism to perform timing analysis on the timing feature sequences, identifying timing-sensitive intervals in the protocol interaction process; the timing attention masking mechanism employs a sliding window mechanism, combined with Transformer positional encoding to calculate the contribution of each time window to the protocol execution success rate, marking windows with a contribution greater than a threshold as timing-sensitive intervals; S4, constructing a dynamic protocol knowledge base based on the comprehensive semantic deviation feature vector and timing-sensitive intervals; S5, generating an operation and maintenance strategy for the baseboard management controller device to be upgraded based on the dynamic protocol knowledge base.

[0009] Furthermore, the Redfish protocol response messages are structured text data in JSON format; the timing feature sequence includes the instruction response delay time sequence during the protocol interaction process, the start and end timestamps of the power policy negotiation window, and the time interval sequence of heartbeat packet transmission.

[0010] The Redfish protocol response message refers to the JSON-formatted structured text data returned by the Baseboard Management Controller (BMC) after receiving a Redfish API request from the Operations and Maintenance Management System. This message conforms to the resource model defined by the DMTF Redfish standard specification and is transmitted via HTTP / HTTPS using a RESTful architecture. The message body contains standard fields (such as metadata fields like @odata.type, @odata.id, Id, and Name) and resource attribute fields (such as Version, Updateable, and Status for FirmwareInventory resources). Additionally, under the OEM extension node, it includes vendor-specific fields (such as HP's "HPE.Firmware Status" and Dell's "Dell.Update Service.Job Queue").

[0011] The instruction response delay time series refers to the ordered numerical sequence of time intervals between the operation and maintenance management system sending Redfish API instructions (such as GET / redfish / v1 / Update Service / Firmware Inventory, POST / redfish / v1 / Update Service / Actions / Simple Update, etc.) and receiving the response message returned by the BMC during the firmware upgrade protocol interaction process.

[0012] The power policy negotiation window refers to the time interval during the protocol handshake phase of the firmware upgrade process, during which the BMC and the operations and maintenance management system negotiate and communicate to determine the power management policy (such as whether hot upgrades are allowed, whether a forced reboot is required, and the power state retention mode) during the upgrade process. This window is defined by a start and end timestamp, and its duration varies depending on the vendor's implementation. For example, HP iLO v2.3 requires this window to last 300±50ms, during which the BMC sends power state query commands and waits for policy confirmation.

[0013] Furthermore, the dynamic protocol knowledge base records the protocol state transition modes of each manufacturer's baseboard management controller at each stage of firmware upgrade; the protocol state transition modes include the trigger time, instruction response delay range, and stage duration of the protocol handshake stage, firmware transmission stage, verification confirmation stage, and restart synchronization stage.

[0014] Further, in S2, the comprehensive semantic deviation feature vector between the vendor's proprietary extended protocol and the standard protocol is extracted, including: constructing protocol modality sample pairs and temporal modality sample pairs; using a BERT-based protocol encoder to perform hierarchical feature extraction on the protocol modality sample pairs to obtain differential feature vectors; using a temporal convolutional network encoder to perform multi-scale temporal pattern extraction on the temporal modality sample pairs to obtain temporal behavior feature vectors; constructing a total loss function including intra-protocol modal contrast loss, intra-temporal modal contrast loss, and cross-modal alignment loss to optimize the parameters of the protocol encoder and the temporal convolutional encoder; and concatenating the differential feature vectors and temporal behavior feature vectors through channels to obtain the comprehensive semantic deviation feature vector.

[0015] The dual-stream self-supervised contrastive learning algorithm refers to a deep learning framework that simultaneously processes two heterogeneous data modalities (protocol text modality and temporal numerical modality). In this application, it automatically learns discriminative feature representations of data by constructing positive and negative sample pairs without manual annotation. "Dual-stream" refers to the algorithm containing two parallel feature extraction paths: a protocol stream encoder (based on the BERT architecture) specifically processes the textual semantic features of Redfish protocol response messages, and a temporal stream encoder (based on a temporal convolutional network) specifically processes the time-series patterns of temporal feature sequences. "Self-supervised" means that the algorithm automatically generates training signals by constructing sample triples (anchor sample, positive sample, and negative sample) without external manual annotation. Positive samples are obtained from anchor samples through data augmentation or subset extraction, and negative samples are extracted from samples of different categories.

[0016] The differential feature vector (DMV) is a high-dimensional numerical vector extracted by the protocol stream encoder from Redfish protocol response messages, quantifying the semantic differences between vendor-specific extended protocols and the DMTF standard protocol. This vector is calculated through a contrastive learning mechanism: complete protocol messages from the same vendor (including standard and OEM fields) are encoded as anchor samples into a vector. The subset of the message containing only DMTF standard fields is encoded as a positive sample into a vector. The difference vector between two vectors This is the differential feature vector, where each dimension of the vector reflects the offset contribution of the private extended field to the protocol semantic space.

[0017] A timing behavior feature vector is a high-dimensional numerical vector extracted by a timing stream encoder from timing feature sequences (including command response delay time sequences, power policy negotiation window timestamps, heartbeat packet transmission interval sequences, etc.) that describes the timing behavior patterns of vendor-specific protocol interactions. This vector captures the dynamic features of timing data at different time scales, such as periodicity, abrupt changes, and trends, through the multi-scale convolutional kernels of a temporal convolutional network (TCN). For example, HP devices exhibit a stable window pattern of 300±50ms during the power policy negotiation phase, while Dell devices exhibit a heartbeat packet pattern with a 2-second cycle during the firmware transmission phase. These vendor-specific timing patterns are encoded into different dimensions of the vector.

[0018] Furthermore, the protocol modality sample pairs are constructed, including: using Redfish protocol response messages from the same vendor and firmware version as anchor samples; extracting a subset of standard fields that meet the DMTF Redfish specification from the anchor samples as positive samples; extracting private extended fields under the OEM node from Redfish protocol response messages collected from different vendors' board management controllers as negative samples; and constructing a protocol modality sample triplet, including anchor samples, positive samples, and negative samples, as a protocol modality sample pair.

[0019] Anchor samples refer to the original sample data used as a reference benchmark during the construction of sample triples in the contrastive learning algorithm. The algorithm learns discriminative feature representations by calculating the feature distances between anchor samples and positive and negative samples. In the construction of protocol modality sample pairs, anchor samples specifically refer to complete Redfish protocol response messages collected from the same manufacturer and firmware version of the baseboard management controller. These messages are presented in JSON format and contain a complete data structure, including both common fields defined according to the DMTF Redfish standard specification (such as metadata and resource attributes like @odata.type, @odata.id, Id, Name, Version, Status, etc.) and proprietary custom fields added by the manufacturer under the OEM extension node (such as HPE.Firmware Status for HP and Dell.Update Service.Job Queue for Dell).

[0020] The Oem node refers to a special attribute node in the JSON data structure of the Redfish protocol response message, which is specifically reserved by the DMTF Redfish standard specification to allow equipment manufacturers (OEMs) to add manufacturer-private extended fields.

[0021] Furthermore, the construction of timing modal sample pairs includes: using timing feature sequences of the same manufacturer and firmware version as anchor samples; applying data augmentation operations to the anchor samples to generate positive samples, wherein the data augmentation operations randomly stretch or compress local segments of the timing sequence on the time axis; extracting timing feature sequences of different manufacturers' board management controllers as negative samples; and constructing timing modal sample triples, including anchor samples, positive samples, and negative samples, as timing modal sample pairs.

[0022] Furthermore, the total loss function is expressed as follows: Where α, β, and γ are weighting coefficients; The intramodal contrastive loss function is used within the protocol. The intramodal contrast loss function; This is the cross-modal alignment loss function.

[0023] Furthermore, in S3, a temporal analysis of the temporal feature sequence is performed using a temporal attention masking mechanism, including: setting a sliding window for the temporal feature sequence; embedding the positional information of the temporal feature sequence within the sliding window using Transformer positional encoding to obtain position-enhanced temporal features; using the position-enhanced temporal features as input, calculating the attention weight matrix for each sliding window using a multi-head attention mechanism; calculating the importance score of each sliding window based on the attention weight matrix, and marking sliding windows with scores greater than a threshold as temporally sensitive intervals.

[0024] Furthermore, in S4, a dynamic protocol knowledge base is constructed, including: using the K-Means clustering algorithm to cluster the comprehensive semantic deviation feature vectors, grouping the baseboard management controllers of the same manufacturer and firmware version into one category to obtain manufacturer device clusters; extracting the cluster center vector of each manufacturer device cluster as the protocol semantic fingerprint of the corresponding manufacturer; statistically analyzing the time distribution of the time-sensitive intervals of each manufacturer device cluster during the protocol interaction process as the time-sensitive interval statistical features of the corresponding manufacturer; constructing the protocol state transition mode of each manufacturer's baseboard management controller at each stage of firmware upgrade based on the protocol semantic fingerprint and the time-sensitive interval statistical features; and storing the protocol state transition mode in the dynamic protocol knowledge base.

[0025] Among them, the protocol semantic fingerprint refers to the cluster center vector extracted from the comprehensive semantic deviation feature vectors of multiple baseboard management controllers of the same manufacturer and firmware version through the K-Means clustering algorithm. This vector serves as a unique identifier for the manufacturer's equipment at the protocol implementation level.

[0026] The statistical features of time-sensitive intervals refer to the set of descriptive features obtained by statistically analyzing all time-sensitive intervals identified in step S3 on the complete timeline of firmware upgrade protocol interaction for each manufacturer's device cluster.

[0027] Another aspect of this application provides a data center operation and maintenance management system for implementing a data center operation and maintenance management method of this application.

[0028] Compared to existing technologies, the advantages of this application are:

[0029] This application utilizes a dual-stream self-supervised contrastive learning algorithm to fuse Redfish protocol text semantic features and temporal behavioral features, enabling automatic extraction of comprehensive semantic deviation feature vectors between vendor-specific extended protocols and standard protocols without manual annotation. Compared to existing technologies that rely on manual maintenance of vendor-specific configuration rules, this method uses the K-Means clustering algorithm to automatically cluster devices from the same vendor and with the same firmware version, extracting protocol semantic fingerprints and constructing a dynamic protocol knowledge base. This knowledge base records the refined protocol state transition patterns of each vendor at each stage of firmware upgrades (protocol handshake, firmware transmission, verification confirmation, and reboot synchronization), including key parameters such as trigger time points, command response latency ranges, and stage durations. This automated knowledge accumulation mechanism allows the system to quickly adapt to new vendor devices and new firmware versions without manual testing or rule writing, significantly reducing maintenance costs for operations personnel and ensuring the knowledge base is continuously updated and optimized with device firmware iterations.

[0030] This application employs a temporal attention masking mechanism combined with Transformer positional encoding. It uses a sliding window approach to perform multi-head attention analysis on the temporal feature sequences of protocol interactions, automatically identifying time-sensitive intervals that affect the success rate of firmware upgrades. Compared to existing technologies that use fixed timeout thresholds, leading to "premature timeout false alarms" or "incomplete upgrades due to insufficient waiting," this method generates differentiated operation and maintenance strategies for the devices to be upgraded based on vendor-specific protocol state transition patterns stored in a dynamic protocol knowledge base. For HP iLO devices, it automatically configures a power policy negotiation window of 300±50ms; for Dell iDRAC devices, it sets a heartbeat packet sending strategy with an interval of ≤2 seconds during the firmware transmission phase; and for Inspur BMC, it configures a response timeout threshold of over 1200ms during the verification phase. Simultaneously, within the identified time-sensitive intervals, the monitoring sampling frequency is increased (device status is collected every 5 seconds), enabling real-time early warning for critical time windows. This adaptive strategy generation mechanism effectively solves the state machine offset problem caused by temporal differences in protocols across multiple vendors, significantly improving the success rate and automation level of firmware upgrades in heterogeneous environments. Attached Figure Description

[0031] This application will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0032] Figure 1 This is an exemplary flowchart illustrating a data center operation and maintenance management method according to some embodiments of this application;

[0033] Figure 2 This is an exemplary flowchart illustrating the construction of protocol modality sample pairs according to some embodiments of this application;

[0034] Figure 3 This is an exemplary flowchart illustrating the construction of temporal modality sample pairs according to some embodiments of this application;

[0035] Figure 4 This is an exemplary flowchart illustrating the construction of a comprehensive semantic deviation feature vector according to some embodiments of this application;

[0036] Figure 5 This is an exemplary flowchart illustrating the construction of a dynamic protocol knowledge base according to some embodiments of this application. Detailed Implementation

[0037] The methods and systems provided in the embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0038] like Figure 1As shown, Redfish protocol response messages and timing feature sequences from multi-vendor board management controllers were collected. The Redfish protocol response messages were structured text data in JSON format. The timing feature sequences included the command response delay time sequence with a sampling frequency of 100Hz during protocol interaction, the start and end timestamps of the power policy negotiation window, and the time interval sequence of heartbeat packet transmission. A dual-stream self-supervised contrastive learning algorithm was used to perform multimodal fusion analysis on the collected Redfish protocol response messages and timing feature sequences, extracting a comprehensive semantic deviation feature vector between vendor-specific extended protocols and the standard protocol. The dual-stream self-supervised contrastive learning algorithm included a protocol stream encoder and a timing stream encoder. The protocol stream encoder extracted protocol semantic features based on the BERT architecture, while the timing stream encoder extracted timing behavior features based on a temporal convolutional network. The two feature streams were fused using a cross-modal alignment loss function. A temporal attention masking mechanism was used to perform temporal analysis on the collected timing feature sequences to identify timing-sensitive areas during protocol interaction. In this process, the timing attention masking mechanism adopts a sampling method with a sliding window step size of 10 milliseconds. Combined with Transformer position encoding, it calculates the contribution of each time window to the protocol execution success rate and marks windows with a contribution greater than 0.3 as timing-sensitive intervals. Based on the extracted comprehensive semantic deviation feature vector and the identified timing-sensitive intervals, a dynamic protocol knowledge base is constructed. The dynamic protocol knowledge base records the protocol state transition modes of each manufacturer's baseboard management controller in each stage of firmware upgrade. The protocol state transition modes include the trigger time points, command response delay ranges, and stage durations of the protocol handshake stage, firmware transmission stage, verification confirmation stage, and restart synchronization stage. Based on the constructed dynamic protocol knowledge base, an adaptive operation and maintenance strategy is generated for the baseboard management controller device to be upgraded. The operation and maintenance strategy includes a protocol compatibility processing scheme determined based on the protocol semantic feature components in the extracted comprehensive semantic deviation feature vector, a power consumption peak time predicted based on the timing behavior feature components, and a command transmission timing adjusted based on the identified timing-sensitive intervals.

[0039] Specifically, Redfish protocol response messages include: Standard fields: Common fields conforming to the DMTF Redfish specification, including @odata.type, @odata.id, Power State, BIOS Version, etc.

[0040] Private extended fields: Vendor-defined fields located under the Oem node, including the Oem.Hp.iLO Self Test Results field for HP board management controller, the Oem.Dell.Dell Attributes field for Dell board management controller, and the Oem.Inspur.BMC Version field for Inspur board management controller.

[0041] The timing feature sequence includes: instruction response delay time sequence: recording the time interval from sending a Redfish request to receiving a response, with no less than 50 sampling points and a sampling frequency of 100Hz; start and end timestamps of the power policy negotiation window: recording the start and end times of the power management policy negotiation process during the protocol handshake phase, with timestamp accuracy at the millisecond level; and heartbeat packet transmission time interval sequence: recording the time interval at which the baseboard management controller actively sends heartbeat packets, with the sampling duration covering the entire firmware upgrade cycle.

[0042] Specifically, the methods for recording protocol state transition modes include:

[0043] The state transition mode of the protocol handshake phase: record the trigger time points of Session establishment, Authentication, and Capability negotiation, the offset of the trigger time point relative to the firmware upgrade task start time, the minimum, maximum, and median values ​​of the command response delay in this phase, and the duration of this phase from start to end.

[0044] State transition mode during firmware transmission: Record the start time of firmware image file chunk upload, record the instruction response delay range of each chunk transmission, record the total duration of this stage and the fluctuation range of transmission rate.

[0045] State transition mode of the verification and confirmation phase: Record the start time of firmware integrity verification, record the instruction response delay range of each step of checksum calculation, hash verification, and digital signature verification, and record the duration of this phase.

[0046] State transition mode during restart synchronization phase: Record the trigger time of baseboard management controller restart, record the time sequence of firmware loading, state recovery, and heartbeat packet reconstruction after restart, and record the duration of this phase from restart to successful session reconnection.

[0047] The index structure of the dynamic protocol knowledge base includes:

[0048] Primary Index: Manufacturer Model Identifier, including manufacturer types such as HP iLO, Dell iDRAC, and Inspur BMC;

[0049] Secondary index: Firmware version number, using a three-segment format of major version number.minor version number.revision number;

[0050] Level 3 Index: Protocol state transition phase identifiers, including four phases: handshake, transmission, verification, and restart;

[0051] Data items: trigger time, delay range, duration of each stage, and the comprehensive semantic deviation feature vector of the firmware version of the manufacturer extracted in step S2.

[0052] Specifically, in step S2, a two-stream self-supervised contrastive learning algorithm is used to extract the comprehensive semantic deviation feature vector between the vendor's proprietary extension protocol and the standard protocol, including the following steps:

[0053] S21, Construct hierarchical sample pairs:

[0054] like Figure 2 As shown, firstly, a protocol modal sample pair is constructed, and the Redfish protocol response messages of the same manufacturer and firmware version collected in step S1 are used as anchor samples, denoted as sample A_protocol;

[0055] Extract a subset of standard fields that meet the DMTFRedfish specification from sample A_protocol. The subset of standard fields includes fields such as @odata.type, @odata.id, Power State, and BIOS Version. This subset of standard fields is taken as a positive sample and denoted as sample P_protocol.

[0056] Private extended fields under the Oem node were extracted from Redfish protocol response messages collected from different manufacturers' board management controllers. These private extended fields include Oem.Hp.iLOSelfTest Results, Oem.Dell.DellAttributes, and Oem.Inspur.BMC Version. These private extended fields from different manufacturers were used as negative samples and denoted as the sample set. ;

[0057] Constructing protocol modal sample triples ,in, For anchor point samples, To and Homologous positive samples This is a set of negative samples from different manufacturers;

[0058] like Figure 3 As shown, time-series modal sample pairs are constructed, and the time-series feature sequences of the same manufacturer and firmware version collected in step S1 are used as anchor samples, denoted as samples. The timing feature sequence includes a command response delay time sequence of length T, the start and end timestamps of the power policy negotiation window, and the time interval sequence of heartbeat packet transmission.

[0059] For anchor point samples Apply data augmentation operations to generate positive samples The specific method for data augmentation is as follows: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] The timeline is randomly divided into M segments, where M ranges from 3 to 7. For each segment, a stretching or compression operation is randomly selected, with a stretching or compression ratio ranging from 0.8 to 1.2. The stretching operation increases the number of sampling points in the segment through linear interpolation, while the compression operation decreases the number of sampling points in the segment through mean pooling. The processed M segments are then reassembled to obtain the positive samples. ;

[0060] Extracting data from timing feature sequences collected from different manufacturers' board management controllers. Time segments of equal length are used as a negative sample set. ;

[0061] Constructing temporal modal sample triples ,in, For anchor point samples, For positive samples enhanced by time warp, This is a set of negative samples from different manufacturers;

[0062] S22, Parallel processing of dual-stream encoder:

[0063] The protocol stream encoder's processing flow is as follows: The protocol mode sample triplet constructed in step S21 is processed... Anchor point samples in The input is a BERT-based protocol encoder, which contains 12 Transformer encoder layers. Each Transformer layer contains a multi-head self-attention module and a feedforward neural network module. The hidden layer dimension is 768, and the number of attention heads is 12.

[0064] A first attention mechanism module is inserted after the 6th Transformer layer of the protocol encoder. This first attention mechanism module processes the samples... Self-attention calculation is performed on the standard fields in the matrix. The features output by the first attention mechanism are reduced to 256 dimensions by a fully connected layer to generate the baseline feature vector z_std_sem.

[0065] A second attention mechanism module is inserted after the 12th Transformer layer of the protocol encoder. This second attention mechanism module computes samples. The cross-attention weights of the private extended fields relative to the standard fields under the Oem node are used. The features output by the second attention mechanism are reduced to 256 dimensions by a fully connected layer to generate a difference feature vector. ;

[0066] Baseline feature vector Capture the general semantic information of standard protocol fields, differential feature vectors Capture the semantic deviation information of private extended fields relative to standard fields, and the two together constitute the vendor characteristic representation at the protocol semantic level;

[0067] Processing flow of a timing stream encoder:

[0068] The time-series modal sample triplet constructed in step S21 Anchor point samples in The input is a temporal convolutional network encoder, which contains three parallel dilated convolutional branches;

[0069] The first branch uses a one-dimensional convolution kernel with a kernel size of 10 milliseconds and an dilation rate of 1, corresponding to one sampling point at a sampling frequency of 100Hz, at the anchor point sample. Convolution operations are performed on the time sequence to extract short-term jitter patterns in the time sequence feature sequence. The short-term jitter patterns reflect millisecond-level latency fluctuations during protocol interaction. The first branch outputs a feature map with 64 channels.

[0070] The second branch uses a one-dimensional convolution kernel with a kernel size of 50 milliseconds and an dilation rate of 5, corresponding to 5 sampling points at a sampling frequency of 100Hz, at the anchor point sample. Convolution operations are performed on the time sequence to extract the intermediate negotiation window pattern in the time sequence feature sequence. The intermediate negotiation window pattern reflects the time sequence structure at tens of milliseconds during processes such as power strategy negotiation. The number of channels in the output feature map of the second branch is 64.

[0071] The third branch uses a one-dimensional convolution kernel with a kernel size of 200 milliseconds and an dilation rate of 20, corresponding to 20 sampling points at a sampling frequency of 100Hz, at the anchor point sample. Convolution operations are performed on the time sequence to extract the long-term heartbeat rhythm pattern in the time feature sequence. The long-term heartbeat rhythm pattern reflects the periodic pattern of the heartbeat packets actively sent by the baseboard management controller. The number of channels in the output feature map of the third branch is 64.

[0072] The feature maps output from the three branches are concatenated along the channel dimension to obtain a fused feature map with 192 channels. The fused feature map contains short-term, medium-term, and long-term temporal pattern information.

[0073] Global average pooling is performed on the fused feature map to compress the time dimension to 1, resulting in a 192-dimensional feature vector.

[0074] The 192-dimensional feature vector is input into a fully connected layer for dimensionality reduction. The fully connected layer contains one hidden layer with 128 neurons and ReLU activation function, and the output layer has 256 neurons, ultimately generating a temporal behavior feature vector. It has 256 dimensions;

[0075] Temporal behavior feature vector The system comprehensively characterizes the vendor's short-term latency fluctuations, medium-term negotiation window characteristics, and long-term heartbeat rhythm characteristics during the protocol interaction process.

[0076] S23, Construct the loss function:

[0077] ;

[0078] Where α, β, and γ are dynamic weight coefficients, which are dynamically adjusted according to the number of training iterations t:

[0079] In the initial training phase, when t ≤ 5000, α = 0.5, β = 0.4, and γ = 0.1 are set. At this time, the protocol intra-modal contrastive loss in the total loss function is... Intramodal contrast loss The dominant approach prioritizes optimizing the feature extraction capabilities of the protocol stream encoder and the temporal stream encoder within their respective modalities.

[0080] In the later stages of training, when t > 5000, adjust α = 0.3, β = 0.3, and γ = 0.4. At this point, the cross-modal alignment loss in the total loss function... Dominant, enhancing the alignment of protocol semantic features and temporal behavioral features in the cross-modal feature space;

[0081] ;in: For the protocol stream encoder in step S22, from the anchor sample The baseline feature vector extracted from the standard fields has a dimension of 256; For the protocol stream encoder in step S22, from the anchor sample The differential feature vector extracted from the private extended field has a dimension of 256; For the protocol stream encoder in step S22, from the negative sample set The private extended field feature vector extracted from the k-th negative sample, where k ranges from 1 to n, and n is the number of negative samples in the batch, usually 64; sim is the cosine similarity function; The temperature hyperparameter is set to 0.07. The temperature hyperparameter is used to control the smoothness of the similarity distribution. A smaller temperature value makes the model more sensitive to differences in similarity.

[0082] ;in: For the timing stream encoder in step S22, from the anchor sample The extracted temporal behavior feature vector has a dimension of 256; For the timing stream encoder in step S22, from the positive samples Extracted temporal behavior feature vectors, positive samples For anchor point samples Generated after time warp enhancement, with a dimension of 256; For the time-series encoder in step S22, from the negative sample set The temporal behavior feature vector extracted from the k-th negative sample, where k ranges from 1 to n, and n is the number of negative samples in the batch, usually 64. The temperature hyperparameter is set to 0.05.

[0083] ;in: For the protocol stream encoder in step S22, from the anchor sample The differential feature vector extracted from the private extended fields has a dimension of 256. The differential feature vector represents the vendor's private extended characteristics at the protocol semantic level. For the timing stream encoder in step S22, from the anchor sample The extracted temporal behavior feature vector has a dimension of 256 and represents the vendor's execution mode characteristics at the temporal behavior level; W is a learnable cross-modal projection matrix with a dimension of 256×256, which projects the temporal behavior feature vector. Projected onto the protocol semantic feature space, the projected vector and the protocol semantic feature vector They reside in the same feature space; ||*||2 is the L2 norm.

[0084] The Adam optimizer is used to simultaneously optimize the parameters of the protocol stream encoder, the temporal stream encoder, and the cross-modal projection matrix W. The learning rate of the Adam optimizer is set to 0.0001. , In each training iteration, a batch of protocol modality sample triplets and temporal modality sample triplets are randomly selected from the sample pairs constructed in step S21, with a batch size of 32. The samples within the batch are then input into the protocol stream encoder and temporal stream encoder of step S22, respectively, to obtain the baseline feature vectors of all samples in that batch. Difference eigenvectors and temporal behavior feature vector According to the total loss function defined in step S23 Calculate the total loss for this batch, where α, β, and γ are dynamically adjusted based on the current iteration number t; calculate the total loss using the backpropagation algorithm. The gradients of the protocol stream encoder parameters, the temporal stream encoder parameters, and the cross-modal projection matrix W are calculated. The Adam optimizer is used to update the protocol stream encoder parameters, the temporal stream encoder parameters, and the cross-modal projection matrix W based on the calculated gradients. The above iterative process is repeated until the total loss L_total converges or the preset maximum number of iterations is reached.

[0085] like Figure 4 As shown in S24, generate the comprehensive semantic deviation feature vector:

[0086] The differential feature vector output by the protocol stream encoder trained in step S23. Timing behavior feature vector output by the timing stream encoder The splicing is performed along the channel dimension, and the splicing formula is as follows: , where [;] denotes vector concatenation operation; differential eigenvectors The dimension is 256, representing the vendor-specific extension characteristics of the protocol semantic layer extracted by the protocol stream encoder in step S22; the temporal behavior feature vector The dimension is 256, representing the vendor execution mode characteristics at the temporal behavior level extracted by the temporal stream encoder in step S22; the concatenated comprehensive semantic deviation feature vector is obtained. The dimension is 512, of which the first 256 dimensions correspond to The last 256 dimensions represent vendor-specific extensions at the semantic level of the protocol, including semantic differences between vendor-specific extension fields under OEM nodes and standard fields; It characterizes the vendor execution mode characteristics at the time sequence behavior level, including short-term latency jitter mode, medium-term negotiation window mode, and long-term heartbeat rhythm mode.

[0087] Specifically, in step S3, a temporal attention masking mechanism is used to perform temporal analysis on the temporal feature sequence to identify temporally sensitive intervals in the protocol interaction process. This includes: setting a sliding window for the temporal feature sequence, with a window length of 128 time steps and a sliding step size of 64 time steps; applying Transformer position encoding to the temporal feature sequence within the sliding window, where the Transformer position encoding uses a sine-cosine function to calculate the position embedding vector for each time step; calculating the attention weight matrix for each sliding window using a multi-head self-attention mechanism, which contains 8 attention heads, each with a dimension of 64; calculating the contribution of each sliding window to the protocol execution success rate based on the attention weight matrix, where the contribution is obtained by summing and normalizing the attention weight matrix along the time dimension; marking sliding windows with a contribution greater than a preset threshold as temporally sensitive intervals, where the preset threshold is set to 0.75; and the temporally sensitive intervals correspond to key time periods in the protocol interaction process, including the period of high packet loss during the firmware transmission phase, the timeout retransmission period during the verification and confirmation phase, and the service interruption period during the restart and synchronization phase.

[0088] Specifically, such as Figure 5As shown in step S4, a dynamic protocol knowledge base is constructed based on the comprehensive semantic deviation feature vector and the time-sensitive interval. This includes: using the K-Means clustering algorithm to cluster the comprehensive semantic deviation feature vector, grouping the baseboard management controllers of the same manufacturer and firmware version into one category to obtain manufacturer device clusters; extracting the cluster center vector of each manufacturer device cluster as the protocol semantic fingerprint of the corresponding manufacturer; statistically analyzing the time distribution of the time-sensitive intervals of each manufacturer device cluster during the protocol interaction process as the time-sensitive interval statistical features of the corresponding manufacturer; constructing the protocol state transition mode of each manufacturer's baseboard management controller at each stage of firmware upgrade based on the protocol semantic fingerprint and the time-sensitive interval statistical features; and storing the protocol state transition mode in the dynamic protocol knowledge base.

[0089] Specifically, in step S5, an operation and maintenance strategy for the baseboard management controller device to be upgraded is generated based on the dynamic protocol knowledge base. This includes: collecting the current Redfish protocol response message and the current timing feature sequence of the baseboard management controller device to be upgraded; inputting the current Redfish protocol response message into a trained protocol stream encoder to extract the protocol semantic feature vector of the device to be upgraded; inputting the current timing feature sequence into a trained timing stream encoder to extract the timing behavior feature vector of the device to be upgraded; performing a vector concatenation operation on the protocol semantic feature vector and the timing behavior feature vector to generate a comprehensive feature vector of the device to be upgraded; traversing all vendor device clusters stored in the dynamic protocol knowledge base, calculating the cosine similarity value between the comprehensive feature vector of the device to be upgraded and the protocol semantic fingerprint of each vendor device cluster, and marking the vendor device cluster with the largest cosine similarity value as the matching cluster.

[0090] Read the protocol state transition mode data records associated with the matching cluster from the dynamic protocol knowledge base. The protocol state transition mode data records include the trigger time point value, upper and lower limit value of the instruction response delay range, and the duration value of each stage for the protocol handshake stage, firmware transmission stage, verification confirmation stage, and restart synchronization stage.

[0091] Read the time-sensitive interval statistical feature data records associated with matching clusters from the dynamic protocol knowledge base, parse the start and end timestamps of the time windows marked with high importance scores in the time-sensitive interval statistical feature data records, and determine the time window as the time window that needs to be monitored in the firmware upgrade process of the device to be upgraded.

[0092] Based on the stage duration values ​​in the protocol state transition mode data records, timeout thresholds are set for each upgrade stage. The timeout threshold is equal to the stage duration value multiplied by a safety factor of 1.5. Based on the upper and lower limits of the command response delay range, a command retry policy is set. When the delay of a single command response exceeds the upper limit of the delay range, the command is resent, with a maximum of 3 retries. Based on the start and end timestamps of the time windows that require key monitoring, real-time monitoring alarm rules are configured. Device status data is collected every 5 seconds within the monitoring time window, and an alarm is triggered immediately when an abnormal status is detected. The timeout threshold configuration, command retry policy, and real-time monitoring alarm rules are combined to generate the operation and maintenance policy for the device to be upgraded.

[0093] The foregoing illustrative description of the present application and its embodiments is not restrictive and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. The accompanying drawings are only one embodiment of the present application, and the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present application, such designs should fall within the scope of protection of this application. Furthermore, the word "comprising" does not exclude other elements or steps, and the word "a" preceding an element does not exclude the inclusion of "a plurality" of that element. Terms such as "first," "second," etc., are used to indicate names and do not indicate any specific order.

Claims

1. A method for operation and maintenance management of a data center, characterized by, The method comprises the following steps: S1, collecting Redfish protocol response messages and time sequence feature sequences of multi-vendor baseboard management controllers; S2, performing multi-modal fusion analysis on the Redfish protocol response messages and the time sequence feature sequences by using a dual-flow self-supervised contrast learning algorithm to extract comprehensive semantic deviation feature vectors of vendor private extension protocols and standard protocols; wherein the dual-flow self-supervised contrast learning algorithm comprises a protocol flow encoder and a time sequence flow encoder, the protocol flow encoder extracts protocol semantic features based on a BERT architecture, and the time sequence flow encoder extracts time sequence behavior features based on a time convolution network; S3, performing time sequence analysis on the time sequence feature sequences by using a time sequence attention mask mechanism to identify time sequence sensitive intervals in a protocol interaction process; the time sequence attention mask mechanism adopts a sliding window mechanism, combines with Transformer position encoding to calculate the contribution of each time window to the protocol execution success rate, and marks the windows with a contribution greater than a threshold as time sequence sensitive intervals; S4, constructing a dynamic protocol knowledge base according to the comprehensive semantic deviation feature vectors and the time sequence sensitive intervals; S5, generating an operation and maintenance strategy for a baseboard management controller device to be upgraded according to the dynamic protocol knowledge base.

2. The operation and maintenance management method of the data center according to claim 1, wherein: the Redfish protocol response message is structured text data in JSON format; the time sequence feature sequence comprises an instruction response delay time sequence in a protocol interaction process, start and end time stamps of a power strategy negotiation window, and a time interval sequence of heartbeat packet sending.

3. The operation and maintenance management method of the data center according to claim 2, wherein: the dynamic protocol knowledge base records protocol state transition modes of each vendor baseboard management controller at each stage of firmware upgrade; the protocol state transition mode comprises trigger time points, instruction response delay ranges and stage duration times of a protocol handshake stage, a firmware transmission stage, a verification and confirmation stage and a restart synchronization stage.

4. The operation and maintenance management method of the data center according to claim 3, wherein: S2, extracting the comprehensive semantic deviation feature vectors of the vendor private extension protocols and the standard protocols comprises: constructing protocol modal sample pairs and time sequence modal sample pairs; extracting hierarchical features of the protocol modal sample pairs by using a BERT-based protocol encoder to obtain a difference feature vector; extracting multi-scale time sequence patterns of the time sequence modal sample pairs by using a time convolution network encoder to obtain a time sequence behavior feature vector; constructing a total loss function comprising a protocol modal intra-contrast loss, a time sequence modal intra-contrast loss and a cross-modal alignment loss for optimizing parameters of the protocol encoder and the time convolution encoder; channel splicing the difference feature vector and the time sequence behavior feature vector to obtain the comprehensive semantic deviation feature vector.

5. The operation and maintenance management method of the data center according to claim 4, wherein: constructing the protocol modal sample pairs comprises: taking Redfish protocol response messages of the same firmware version of the same vendor as anchor samples; extracting a standard field subset meeting the DMTF Redfish specification from the anchor samples as positive samples; The private extension field under the Oem node in the Redfish protocol response message collected from the baseboard management controller of different manufacturers is extracted as a negative sample. A protocol modal sample triple is constructed, including an anchor sample, a positive sample, and a negative sample, as a protocol modal sample pair.

6. The operation and maintenance management method of the data center according to claim 4, characterized in that: A time sequence modal sample pair is constructed, including: The time sequence feature sequence of the same manufacturer and the same firmware version is taken as an anchor sample; A data enhancement operation is applied to the anchor sample to generate a positive sample, and the data enhancement operation is to randomly stretch or compress the local paragraph of the time sequence in the time axis; The time sequence feature sequence of the baseboard management controller of different manufacturers is extracted as a negative sample; A time sequence modal sample triple is constructed, including an anchor sample, a positive sample, and a negative sample, as a time sequence modal sample pair.

7. The operation and maintenance management method of the data center according to claim 4, characterized in that: The total loss function expression is as follows: ; wherein, α, β, γ are weight coefficients; is a protocol intra-modal contrast loss function; is a time sequence intra-modal contrast loss function; is a cross-modal alignment loss function.

8. The operation and maintenance management method of the data center according to any one of claims 4 to 7, characterized in that: S3, using a time sequence attention mask mechanism to perform time sequence analysis on the time sequence feature sequence, including: Setting a sliding window of the time sequence feature sequence; Using the Transformer position encoding to embed the position information of the time sequence feature sequence in the sliding window to obtain position-enhanced time sequence features; Using a multi-head attention mechanism to calculate an attention weight matrix of each sliding window by taking the position-enhanced time sequence features as input; According to the attention weight matrix, the importance score of each sliding window is calculated, and the sliding window with a score greater than a threshold value is marked as a time sequence sensitive interval.

9. The operation and maintenance management method of the data center according to claim 8, characterized in that: S4, constructing a dynamic protocol knowledge base, including: Using a K-Means clustering algorithm to cluster the comprehensive semantic deviation feature vectors, and clustering the baseboard management controllers of the same manufacturer and the same firmware version into a class to obtain a manufacturer device class cluster; Extracting the cluster center vector of each manufacturer device class cluster as the protocol semantic fingerprint of the corresponding manufacturer; Statistically analyzing the time distribution of the time sequence sensitive interval in the protocol interaction process of each manufacturer device class cluster as the time sequence sensitive interval statistical characteristics of the corresponding manufacturer; According to the protocol semantic fingerprint and the time sequence sensitive interval statistical characteristics, constructing the protocol state transition mode of the baseboard management controller of each manufacturer in each stage of firmware upgrade; Storing the protocol state transition mode to the dynamic protocol knowledge base.

10. An operation and maintenance management system of a data center, characterized in that, including: At least one processing unit for executing instructions to implement the operation and maintenance management method of the data center according to any one of claims 1 to 9.

Images