Anti-seismic monitoring method and system for server cabinet
By combining multimodal sensors and graph neural networks, the problems of noise interference and misjudgment in seismic monitoring of server cabinets are solved, and highly accurate and timely seismic monitoring is achieved, ensuring the safe and stable operation of the data center.
Patent Information
- Application Number
- CN202510997616.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-19
- Publication Date
- 2025-10-28
AI Technical Summary
Existing server cabinet seismic monitoring methods suffer from high noise interference and misjudgment rate during signal processing, which affects the accuracy of early warning and leads to insufficient timeliness and effectiveness of seismic protection measures.
Multimodal vibration sensors and noise reference sensors are used to synchronously collect signal data, and the signals are processed by combining blind source separation and adaptive noise reduction algorithms. Graph neural networks are used to identify earthquake events, and early warnings or emergency responses are triggered through multi-scale time-frequency decomposition and comprehensive earthquake risk indicator evaluation.
Effectively reduce noise interference, improve signal resolution and warning accuracy, reduce misjudgment rate, ensure timely activation of earthquake protection measures, avoid waste of resources and missed critical warning opportunities, and improve the security and stability of data centers.
Smart Images
Figure CN120846489A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of seismic monitoring technology, specifically to a seismic monitoring method and system for server racks. Background Technology
[0002] The seismic monitoring method for server racks is mainly to ensure the safe and stable operation of data centers and server equipment during earthquakes and other vibration events. By monitoring the vibration in real time, it can provide timely warnings and take protective measures to reduce the risk of equipment damage and data loss.
[0003] However, existing seismic monitoring methods for server racks often have the following technical drawbacks when applied to signal processing and data analysis in data centers:
[0004] 1. Significant signal noise interference:
[0005] Vibration signals acquired by vibration sensors are often accompanied by a significant amount of environmental noise, such as mechanical vibrations generated by the equipment itself and electromagnetic interference. This noise interference can affect the accuracy of the signal, reduce the effective resolution of the data, and make it difficult for subsequent signal processing to accurately separate real seismic vibrations from background noise.
[0006] 2. The false positive rate is relatively high, affecting the accuracy of early warnings:
[0007] Noise interference blurs signal characteristics, making it difficult for signal processing algorithms (such as spectral analysis and wavelet transform) to accurately extract valid seismic signals, easily leading to misjudgments or missed detections. For example, vibrations from equipment operation may be misjudged as seismic vibrations, or seismic early warning signals may fail to be captured in a timely manner. Such misjudgments not only affect the timely activation of seismic protection measures but may also lead to a waste of maintenance resources or the loss of critical early warning opportunities. Summary of the Invention
[0008] To address the shortcomings of existing technologies, this invention provides a method and system for seismic monitoring of server racks, which solves the technical deficiencies mentioned in the background section.
[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for seismic monitoring of server racks, comprising the following steps:
[0010] In advance, an electronic map of the server rack layout within the data center is obtained, and multimodal vibration sensors and noise reference sensors are deployed at key stress points of the racks and in the surrounding environment to simultaneously collect raw vibration signal datasets and environmental noise signal datasets.
[0011] Based on the vibration signal dataset and the environmental noise signal dataset, a set of denoised vibration features is obtained by using a fusion processing of blind source separation and adaptive noise reduction algorithms.
[0012] The noise suppression coefficient Nsc is calculated based on the set of denoised vibration features, and the signal quality threshold Nth is preset. The noise suppression coefficient Nsc is compared with the signal quality threshold Nth to determine the validity of the current vibration data and decide whether to enter the subsequent seismic event identification process.
[0013] After determining the validity of the vibration data, multi-scale time-frequency decomposition techniques, such as wavelet packet decomposition and short-time Fourier transform, are used to perform deep feature extraction on the denoised vibration feature set, and obtain the energy spectrum features characterizing the vibration energy distribution and the typical spectral mode features reflecting the vibration characteristics.
[0014] The vibration energy distribution characteristics and typical spectral pattern characteristics are input into the trained graph neural network seismic classification model, and the seismic event confidence Cev is output.
[0015] Based on the earthquake event confidence level Cev and the vibration energy distribution characteristics, the comprehensive earthquake risk index SRI is calculated, and a risk threshold Tth is pre-set for comparative evaluation, so as to analyze the earthquake safety status of the server rack in real time and trigger corresponding early warning or emergency response commands.
[0016] Preferably, based on the electronic map of the server rack layout, the top, base and middle stress nodes of each server rack are determined, and high-sensitivity triaxial accelerometers and environmental noise sensors are deployed to collect vibration signal data and environmental noise signal data simultaneously.
[0017] A time synchronization protocol is used to correct the data delay of multiple sensors. Low-frequency and high-frequency components that are not related to the seismic frequency band are initially filtered out by a bandpass filter in order to construct initial, multi-channel vibration signal datasets and environmental noise signal datasets, respectively.
[0018] Preferably, based on the denoised vibration feature set, blind source separation is achieved through independent component analysis, and the seismic vibration component Se and noise component Sn are extracted respectively, and Wiener filtering is used sequentially for adaptive noise suppression;
[0019] The noise suppression coefficient Nsc is obtained by correlating the seismic vibration component Se and the noise component Sn.
[0020] When the noise suppression coefficient Nsc is less than the signal quality threshold Nth, the ICA and Wiener filter parameters are automatically adjusted to optimize the noise reduction effect, and the process is iterated until the noise suppression coefficient Nsc is greater than or equal to the signal quality threshold Nth.
[0021] Preferably, when the noise suppression coefficient Nsc is greater than or equal to the signal quality threshold Nth, the multi-scale time-frequency decomposition module is started to perform three-level wavelet packet decomposition and short-time Fourier transform on the denoised vibration feature set, and extract the sub-band energy Ev[i] and the main frequency component Fv[i] of each scale.
[0022] Preferably, the internal environmental operating parameters of the cabinet are collected synchronously, including CPU temperature Tcpu, fan speed Rfan, and internal humidity Hum. The above operating parameters are fused with the extracted energy spectrum features and spectral pattern features to construct a multi-dimensional feature dataset to enhance the contextual information and model robustness of seismic event discrimination.
[0023] Preferably, the seismic classification model is a graph neural network, in which each sensor node of the server rack constitutes a graph vertex, the node features are vibration energy distribution features and typical spectral pattern features, and the graph edges are defined as physical adjacency relationships; through multi-layer convolution and attention mechanisms, the spatiotemporal correlation is learned in real time, and the seismic event confidence score Cev is output for classifying real earthquake events and equipment self-vibration.
[0024] Preferably, based on the vibration energy distribution characteristics, typical spectral pattern characteristics and seismic event confidence Cev, historical benchmark energy Eref and spectral template constant Fref[i] are extracted, and α, β and γ are set as weighting coefficients to calculate the comprehensive seismic risk index SRI.
[0025] Preferably, a seismic risk threshold Tth is preset. When the comprehensive seismic risk index SRI is greater than or equal to the seismic risk threshold Tth, the system issues a high-risk warning command and automatically executes the following emergency response:
[0026] Trigger the enhanced mode of the cabinet vibration isolation support, push an alarm to the upper-level monitoring platform, and initiate the UPS safety shutdown procedure;
[0027] When the comprehensive seismic risk index SRI is less than the seismic risk threshold Tth, the routine monitoring status is maintained.
[0028] Preferably, the system records key parameters Nsc, vibration energy distribution characteristics, typical spectral pattern characteristics, seismic event confidence level Cev, and comprehensive seismic risk index SRI in real time during the monitoring process, and asynchronously uploads the monitoring logs and model outputs to the cloud database through the edge gateway to support subsequent big data analysis and online model retraining.
[0029] Preferably, based on cloud-based big data analysis results, the monitoring plan is dynamically adjusted, including: optimizing sensor deployment density and location, fine-tuning ICA and Wiener filter parameters online, regularly updating graph neural network model weights, enhancing the performance of rack vibration isolation devices, and regularly organizing seismic emergency drills for maintenance personnel, so as to continuously improve the seismic monitoring capabilities of data center server racks.
[0030] This invention provides a method and system for seismic monitoring of server racks. It has the following beneficial effects:
[0031] (1) The seismic monitoring method and system for server racks adopts multimodal vibration sensors and noise reference sensors to simultaneously collect original vibration signal datasets and environmental noise signal datasets, and combines blind source separation and adaptive noise reduction algorithms to effectively process and reduce interference in environmental noise, such as mechanical vibration and electromagnetic interference generated by the operation of the equipment itself; these processing improve the effective resolution of the signal, so that subsequent signal processing such as multi-scale time-frequency decomposition can more accurately extract vibration energy distribution features Ev[i] and typical spectral mode features Fv[i] from the denoised vibration feature set, thereby enhancing the separation of real earthquake vibration from background noise.
[0032] (2) The seismic monitoring method and system for server racks can effectively learn and identify the vibration energy distribution characteristics and typical spectrum pattern characteristics through the application of graph neural network seismic classification model, and output high-confidence seismic event Cev; combined with the calculated comprehensive seismic risk index SRI, this model evaluates the seismic safety status based on the preset risk threshold Tth, which significantly reduces the misjudgment rate and missed judgment rate caused by noise interference and feature ambiguity, ensures the timely activation of seismic protection measures, and avoids the waste of operation and maintenance resources and the loss of key early warning opportunities caused by misjudgment or missed judgment. Attached Figure Description
[0033] Figure 1 This is a schematic diagram of the method flow of the present invention;
[0034] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation
[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0036] Example 1
[0037] Please see Figure 1 This invention provides a method for seismic monitoring of server racks, comprising the following steps:
[0038] In advance, an electronic map of the server rack layout within the data center is obtained, and multimodal vibration sensors and noise reference sensors are deployed at key stress points of the racks and in the surrounding environment to simultaneously collect raw vibration signal datasets and environmental noise signal datasets.
[0039] Based on the vibration signal dataset and the environmental noise signal dataset, a set of denoised vibration features is obtained by using a fusion processing of blind source separation and adaptive noise reduction algorithms.
[0040] The noise suppression coefficient Nsc is calculated based on the set of denoised vibration features, and the signal quality threshold Nth is preset. The noise suppression coefficient Nsc is compared with the signal quality threshold Nth to determine the validity of the current vibration data and decide whether to enter the subsequent seismic event identification process.
[0041] After determining the validity of the vibration data, multi-scale time-frequency decomposition techniques, such as wavelet packet decomposition and short-time Fourier transform, are used to perform deep feature extraction on the denoised vibration feature set, and obtain the energy spectrum features characterizing the vibration energy distribution and the typical spectral mode features reflecting the vibration characteristics.
[0042] The vibration energy distribution characteristics and typical spectral pattern characteristics are input into the trained graph neural network seismic classification model, and the seismic event confidence Cev is output.
[0043] Based on the earthquake event confidence level Cev and the vibration energy distribution characteristics, the comprehensive earthquake risk index SRI is calculated, and a risk threshold Tth is pre-set for comparative evaluation, so as to analyze the earthquake safety status of the server rack in real time and trigger corresponding early warning or emergency response commands.
[0044] Furthermore, by acquiring an electronic map of the server rack layout within the data center and deploying multimodal vibration sensors and noise reference sensors at key stress points and in the surrounding environment, the original vibration signal dataset and environmental noise signal dataset can be collected simultaneously. This allows for a comprehensive understanding of the real-time dynamics of the racks and the influencing factors of the surrounding environment.
[0045] These data are processed using blind source separation and adaptive noise reduction algorithms to efficiently remove noise interference and improve the purity and reliability of the data. The noise suppression coefficient Nsc is calculated based on the denoised vibration feature set and compared with the pre-set signal quality threshold Nth, which helps to quickly determine the validity of the vibration data and ensure that subsequent processing is based on accurate data.
[0046] By using multi-scale time-frequency decomposition technology to extract deep features from effective data, we can accurately obtain the characteristics of vibration energy distribution and typical spectral patterns, providing detailed feature basis for subsequent data analysis and event identification. When these features are input into a trained graph neural network seismic classification model, the model can output a high-confidence seismic event confidence score Cev, indicating that the model can effectively distinguish between real vibration events and general vibration noise.
[0047] Finally, based on the earthquake event confidence level (Cev) and vibration energy distribution characteristics, a comprehensive seismic risk index (SRI) is calculated and compared with a pre-set risk threshold (Tth). This allows for timely and accurate assessment of the seismic safety status of server racks and triggers corresponding early warning or emergency response measures, enhancing the data center's security management and disaster prevention capabilities. These measures collectively improve data accuracy and response efficiency, ensuring the stable operation and data security of the data center in the event of disasters such as earthquakes.
[0048] Based on the electronic map of the server rack layout, the top, base and middle stress nodes of each server rack are determined, and high-sensitivity triaxial accelerometers and environmental noise sensors are deployed to collect vibration signal data and environmental noise signal data simultaneously.
[0049] A time synchronization protocol is used to correct the data delay of multiple sensors. Low-frequency and high-frequency components that are not related to the seismic frequency band are initially filtered out by a bandpass filter in order to construct initial, multi-channel vibration signal datasets and environmental noise signal datasets, respectively.
[0050] Based on the denoised vibration feature set, blind source separation is achieved through independent component analysis, and the seismic vibration component Se and noise component Sn are extracted respectively. Then, Wiener filtering is used for adaptive noise suppression.
[0051] The noise suppression coefficient Nsc is obtained by correlating the seismic vibration component Se and the noise component Sn.
[0052] When the noise suppression coefficient Nsc is less than the signal quality threshold Nth, the ICA and Wiener filter parameters are automatically adjusted to optimize the noise reduction effect, and the process is iterated until the noise suppression coefficient Nsc is greater than or equal to the signal quality threshold Nth.
[0053] Furthermore, based on the pre-obtained electronic map of the server rack layout within the data center, the locations of key stress points, such as the top, base, and middle of each server rack, are determined. The adequacy of this step is crucial, as thorough preparation directly impacts the quality and efficiency of all subsequent monitoring and analysis. High-sensitivity triaxial accelerometers and environmental noise sensors are deployed at these stress points to simultaneously collect vibration and environmental noise signal data.
[0054] To synchronize data from multiple sensors and ensure the accuracy of data timing, this invention employs a time synchronization protocol, namely the NTP protocol, to correct for delays between data from multiple sensors. Subsequently, the signal is processed through a bandpass filter, which is designed to retain only the frequency band related to seismic vibration. Typically, the seismic signal frequency band is set to 0.1Hz to 50Hz, thereby filtering out high-frequency and low-frequency components unrelated to the seismic frequency band to construct the initial multi-channel vibration signal dataset and environmental noise signal dataset.
[0055] Based on the denoised vibration feature set, independent component analysis is applied to separate the blind sources of the signal, thereby extracting the seismic vibration component Se and the noise component Sn respectively, and then Wiener filtering is used for adaptive noise suppression.
[0056] The parameter settings of Wiener filtering depend on the power spectral density of the signal and noise, and these parameters need to be dynamically adjusted according to real-time data in practical applications;
[0057] The specific formula for calculating the noise suppression coefficient Nsc is as follows:
[0058]
[0059] When the noise suppression coefficient Nsc is greater than or equal to the signal quality threshold Nth, the multi-scale time-frequency decomposition module is started to perform three-level wavelet packet decomposition and short-time Fourier transform on the denoised vibration feature set, and extract the sub-band energy Ev[i] and the main frequency component Fv[i] of each scale.
[0060] Simultaneously collect environmental parameters inside the server rack, including CPU temperature (Tcpu), fan speed (Rfan), and internal humidity (Hum). These parameters are then fused with extracted energy spectrum features and spectral pattern features to construct a multi-dimensional feature dataset, enhancing the contextual information and robustness of the seismic event discrimination model.
[0061] The seismic classification model is a graph neural network. The sensor nodes of the server rack constitute the vertices of the graph. The node features are the vibration energy distribution features and typical spectrum pattern features. The graph edges are defined as physical adjacency relationships. Through multi-layer convolution and attention mechanisms, the spatiotemporal correlation is learned in real time, and the seismic event confidence score Cev is output to classify real earthquake events and equipment self-vibration.
[0062] Based on the vibration energy distribution characteristics, typical spectral pattern characteristics, and seismic event confidence Cev, historical benchmark energy Eref and spectral template constant Fref[i] are extracted, and α, β, and γ are set as weighting coefficients to calculate the comprehensive seismic risk index SRI.
[0063] A pre-set seismic risk threshold Tth is established. When the comprehensive seismic risk index SRI is greater than or equal to the seismic risk threshold Tth, the system issues a high-risk warning command and automatically executes the following emergency response:
[0064] Trigger the enhanced mode of the cabinet vibration isolation support, push an alarm to the upper-level monitoring platform, and initiate the UPS safety shutdown procedure;
[0065] When the comprehensive seismic risk index SRI is less than the seismic risk threshold Tth, the routine monitoring status is maintained.
[0066] Furthermore, when the noise suppression coefficient Nsc is greater than or equal to the signal quality threshold Nth, the steps further include starting the multi-scale time-frequency decomposition module, which first performs three-level wavelet packet decomposition on the denoised vibration feature set to separate vibration signals in different frequency ranges. The wavelet transform details of each level can represent the sub-band energy Ev[i] of the vibration signal.
[0067] Next, a short-time Fourier transform is performed to calculate the dominant frequency component Fv[i] in each time window, where i represents a different frequency band; the energy and dominant frequency of each frequency band are calculated using the following formula:
[0068]
[0069] in, It is the sum of squares of the wavelet transform results of the i-th frequency band, while the main frequency component Fv[i] is located at the point of highest frequency energy based on the STFT results.
[0070] This step also includes synchronously collecting environmental parameters inside the cabinet, specifically CPU temperature (Tcpu) and fan speed (Rfan). These parameters reflect the cabinet's operating status and environmental conditions, serving as auxiliary analysis data to enhance the identification of seismic events. A multi-dimensional monitoring dataset is constructed to integrate vibration characteristics and environmental parameters, providing a more comprehensive perspective for subsequent data analysis and event determination.
[0071] Subsequently, based on the acquired vibration energy distribution characteristics and typical spectral pattern characteristics, the data are input into a trained graph neural network model. This model treats each sensor node in the server rack as a vertex in the graph. The node features include vibration energy and spectral pattern characteristics, and the physical adjacency relationships constitute the graph edges. The model's spatiotemporal correlation learning ability is enhanced through multi-layer convolution and attention mechanisms, and the model outputs the earthquake resistance event confidence score Cev, which effectively distinguishes between real earthquakes and equipment self-vibration.
[0072] Next, by comprehensively utilizing vibration energy distribution characteristics, spectral mode characteristics, and seismic event confidence level (Cev), the comprehensive seismic risk index (SRI) is calculated, as follows:
[0073]
[0074] Where Eref is the historical vibration energy benchmark, Fref[i] is the spectral template constant, and α, β, and γ are weighting coefficients. These weights can be used to adjust the influence of each factor on the risk index.
[0075] This technical solution integrates sophisticated time-frequency analysis technology, environmental monitoring, advanced graph neural network learning, and dynamic risk assessment based on comprehensive indicators to form a comprehensive seismic monitoring system from data acquisition to emergency response. This effectively improves the data center's ability to cope with seismic events such as earthquakes, ensuring system stability and data security.
[0076] The system records key parameters Nsc, vibration energy distribution characteristics, typical spectral pattern characteristics, earthquake event confidence level Cev, and comprehensive seismic risk index SRI in real time during the monitoring process. The monitoring logs and model outputs are asynchronously uploaded to the cloud database through the edge gateway, supporting subsequent big data analysis and online model retraining.
[0077] Furthermore, based on cloud-based big data analysis results, the monitoring plan is dynamically adjusted, including: optimizing sensor deployment density and location, fine-tuning ICA and Wiener filter parameters online, regularly updating graph neural network model weights, enhancing the performance of rack vibration isolation devices, and regularly organizing seismic emergency drills for maintenance personnel to continuously improve the seismic monitoring capabilities of data center server racks.
[0078] After confirming that the noise suppression coefficient Nsc reaches the signal quality threshold Nth, the system enters the data analysis and uploading phase. This step involves real-time recording of key monitoring parameters, including vibration energy distribution characteristics, typical spectral pattern characteristics, seismic event confidence level Cev, and comprehensive seismic risk index SRI. These monitoring logs and model outputs are then asynchronously uploaded to the cloud database via edge gateway technology. This design not only optimizes data storage and processing efficiency but also supports large-scale data analysis and online model retraining, providing the system with continuous learning and gradual optimization capabilities.
[0079] Furthermore, this method dynamically adjusts monitoring strategies based on big data analysis results returned from a cloud database to further improve the efficiency and accuracy of the monitoring system. Specific adjustments include, but are not limited to, optimizing the density and location of sensor deployments, fine-tuning ICA and Wiener filter parameters online, and periodically updating the weights of the graph neural network model. In addition, the system will enhance the performance of the cabinet's vibration isolation devices to more effectively protect hardware equipment. Simultaneously, regular seismic emergency drills will be organized for maintenance personnel to ensure all participants are familiar with emergency procedures, enabling a rapid and effective response in real seismic events.
[0080] Taking all factors into consideration, this technical solution combines the advantages of cloud computing and edge computing, achieving real-time data processing and long-term optimization through deep learning. It provides a comprehensive, dynamic, and continuously improving seismic monitoring system for data center server racks. This system design not only improves the real-time performance and accuracy of monitoring data but also enhances the system's adaptability and early warning capabilities through continuous optimization of monitoring strategies. By implementing these technical features, the security and stability of data centers are improved, effectively preventing and mitigating the potential impact of disasters such as earthquakes on critical equipment. The following is a specific implementation example:
[0081] Example 2
[0082] Please see Figure 2 A seismic monitoring system for server racks, comprising the following modules:
[0083] The layout sensor deployment module pre-acquires an electronic map of the server rack layout within the data center and deploys multimodal vibration sensors and noise reference sensors at key stress nodes of the racks and in the surrounding environment to simultaneously collect raw vibration signal datasets and environmental noise signal datasets.
[0084] The noise fusion processing module, based on the vibration signal dataset and the environmental noise signal dataset, uses blind source separation and adaptive noise reduction algorithms to fuse and process them to obtain a set of denoised vibration features.
[0085] The quality threshold determination module calculates the noise suppression coefficient Nsc based on the denoised vibration feature set and pre-sets the signal quality threshold Nth. It compares the noise suppression coefficient Nsc with the signal quality threshold Nth to determine the validity of the current vibration data and decide whether to proceed to the subsequent seismic event identification process.
[0086] The time-frequency feature extraction module, after determining that the vibration data is valid, performs deep feature extraction on the denoised vibration feature set based on multi-scale time-frequency decomposition to obtain vibration energy distribution features and typical spectral mode features.
[0087] The neural network judgment module inputs the vibration energy distribution characteristics and typical spectral pattern characteristics into the trained graph neural network seismic classification model and outputs the seismic event confidence Cev.
[0088] The risk indicator assessment module calculates the comprehensive seismic risk index SRI based on the seismic event confidence level Cev and vibration energy distribution characteristics, and compares and evaluates the pre-set risk threshold Tth to analyze the seismic safety status of the server rack in real time and trigger corresponding early warning or emergency response commands.
[0089] Furthermore, suppose a data center is affected by a magnitude 6.5 earthquake;
[0090] Monitoring object: Cabinet number RACK-07, that is, 10 three-axis acceleration sensors and 2 noise sensors are deployed;
[0091] Timestamp: 2025-05-24 14:05:32.456;
[0092] Table 1 shows the content of sensor data acquisition and preprocessing;
[0093] Sensor type Location Original signal peak Filtered effective value Accelerometer Top node 0.35 0.28 Accelerometer Base node 0.41 0.33 noise sensor 1m from the server rack 0.18 0.12
[0094] The preprocessing process is: band-pass filtering, that is, 0.1Hz~35Hz;
[0095] Remove the low-frequency vibration of the fan, that is, <0.1Hz;
[0096] High-frequency electromagnetic interference, that is, >100Hz;
[0097] Table 2 shows the content of noise suppression processing and effectiveness determination;
[0098] signal component Energy value Noise suppression coefficient Nsc Signal quality threshold Nth Judgment results Seismic component Se 0.31 0.89 0.85 efficient Noise component Sn 0.08 - - -
[0099] The calculation process is: ;
[0100] Iterative optimization: Initial Nsc = 0.82 < Nth, and it meets the standard after 3 adjustments of Wiener filter parameters
[0101] Table 3 shows the content of multi-scale feature extraction, that is, the energy distribution of three-layer wavelet packet decomposition:
[0102] frequency band 0-5Hz 5-10Hz 10-20Hz 20-35Hz Energy Ev[i] 0.15 0.42 0.28 0.05
[0103] Table 4 shows the content of short-time Fourier transform features:
[0104] The main frequency component Fv[i] 8.2Hz (peak) 16.7Hz (secondary peak)
[0105] Table 5 shows the supplement of operating condition parameters:
[0106] CPU temperature (°C) Fan speed (RPM) 62 8，200
[0107] Perform graph neural network classification;
[0108] Input features:
[0109] Vertex features: [Ev[0:3], Fv[0:1], Tc, Rfan] = [0.15, 0.42, 0.28, 8.2, 62, 8200]
[0110] Graph structure: Construct an adjacency matrix (fully connected with 10 nodes) based on the physical connection of the cabinet
[0111] Table 6 shows the output results:
[0112] Classification results Confidence level Cev Real earthquake events 0.92 Equipment self-vibration (false alarm) 0.05
[0113] Calculation of comprehensive seismic risk indicators;
[0114] Parameter settings:
[0115] Historical baseline: Eref=0.12 (mean value under seismic-free conditions), Fref=[5.0, 12.0];
[0116] Weighting coefficients: α=0.6 (energy), β=0.3 (spectrum), γ=0.1 (confidence level);
[0117] The calculation process is as follows:
[0118] SRI = 0.6 * 7.5 + 0.3 * 3.2 + 0.092 = 4.5 + 0.96 + 0.092 = 5.552 → Normalized to 0.87 (range 0 ~ 1);
[0119] Table 7 shows the triggering and response process for early warnings:
[0120] Risk Indicator SRI Risk threshold Tth determination Response Action 0.87 0.85 High risk 1. Trigger the vibration isolation support's boost mode. 2. Send an ALERT-3 level alarm to the monitoring center. 3. Initiate UPS safety shutdown (delay 10 seconds).
[0121] Data records are synchronized with the cloud;
[0122] Cloud-based optimization feedback;
[0123] Table 8 shows the model update record table:
[0124] Optimization items Before adjustment After adjustment Effect Sensor deployment density 8 per cabinet 12 per cabinet Feature extraction error ↓15% Number of ICA iterations 3 times 1 time (adaptive) Calculation delay ↓200ms Graph Neural Network Weights V1.2 V1.3 False positive rate ↓8% (validation set) Seismic isolation device response threshold SRI>0.85 SRI>0.82 Warning time increased by 1.2 seconds.
[0125] Table 9 is a summary table of key data:
[0126] stage Core parameters Numerical Threshold / Benchmark Noise suppression Nsc 0.89 Nth=0.85 Feature extraction Main frequency energy (5-10Hz) 0.42 Eref=0.12 Event Classification Earthquake confidence level Cev 0.92 - risk assessment SRI 0.87 Tth=0.85 Response results Warning level ALERT-3 -
[0127] This example fully demonstrates the entire process from seismic wave arrival → feature extraction → intelligent decision-making → emergency response. The data tables strictly follow the algorithm logic and parameter definitions in the patent claims, meeting the real-time and accuracy requirements of industrial-grade monitoring systems.
[0128] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
Claims
1. A method for seismic monitoring of server racks, characterized in that: Includes the following steps: In advance, an electronic map of the server rack layout within the data center is obtained, and multimodal vibration sensors and noise reference sensors are deployed at key stress points of the racks and in the surrounding environment to simultaneously collect raw vibration signal datasets and environmental noise signal datasets. Based on the vibration signal dataset and the environmental noise signal dataset, a set of denoised vibration features is obtained by using a fusion processing of blind source separation and adaptive noise reduction algorithms. The noise suppression coefficient Nsc is calculated based on the set of denoised vibration features, and the signal quality threshold Nth is preset. The noise suppression coefficient Nsc is compared with the signal quality threshold Nth to determine the validity of the current vibration data and decide whether to enter the subsequent seismic event identification process. After determining the validity of the vibration data, multi-scale time-frequency decomposition techniques, such as wavelet packet decomposition and short-time Fourier transform, are used to perform deep feature extraction on the denoised vibration feature set, and obtain the energy spectrum features characterizing the vibration energy distribution and the typical spectral mode features reflecting the vibration characteristics. The vibration energy distribution characteristics and typical spectral pattern characteristics are input into the trained graph neural network seismic classification model, and the seismic event confidence Cev is output. Based on the earthquake event confidence level Cev and the vibration energy distribution characteristics, the comprehensive earthquake risk index SRI is calculated, and a risk threshold Tth is pre-set for comparative evaluation, so as to analyze the earthquake safety status of the server rack in real time and trigger corresponding early warning or emergency response commands.
2. The seismic monitoring method for a server rack according to claim 1, characterized in that: Based on the electronic map of the server rack layout, the top, base and middle stress nodes of each server rack are determined, and high-sensitivity triaxial accelerometers and environmental noise sensors are deployed to collect vibration signal data and environmental noise signal data simultaneously. A time synchronization protocol is used to correct the data delay of multiple sensors. Low-frequency and high-frequency components that are not related to the seismic frequency band are initially filtered out by a bandpass filter in order to construct initial, multi-channel vibration signal datasets and environmental noise signal datasets, respectively.
3. The seismic monitoring method for a server rack according to claim 1, characterized in that: Based on the denoised vibration feature set, blind source separation is achieved through independent component analysis, and the seismic vibration component Se and noise component Sn are extracted respectively. Then, Wiener filtering is used for adaptive noise suppression. The noise suppression coefficient Nsc is obtained by correlating the seismic vibration component Se and the noise component Sn. When the noise suppression coefficient Nsc is less than the signal quality threshold Nth, the ICA and Wiener filter parameters are automatically adjusted to optimize the noise reduction effect, and the process is iterated until the noise suppression coefficient Nsc is greater than or equal to the signal quality threshold Nth.
4. The seismic monitoring method for a server rack according to claim 1, characterized in that: When the noise suppression coefficient Nsc is greater than or equal to the signal quality threshold Nth, the multi-scale time-frequency decomposition module is started to perform three-level wavelet packet decomposition and short-time Fourier transform on the denoised vibration feature set, and extract the sub-band energy Ev[i] and the main frequency component Fv[i] of each scale.
5. The seismic monitoring method for a server rack according to claim 1, characterized in that: Simultaneously collect environmental operating parameters inside the cabinet, including CPU temperature (Tcpu), fan speed (Rfan), and humidity (Hum). These operating parameters are then fused with extracted energy spectrum features and spectral pattern features to construct a multi-dimensional feature dataset, thereby enhancing the contextual information and model robustness for seismic event discrimination.
6. The seismic monitoring method for a server rack according to claim 1, characterized in that: The seismic classification model is a graph neural network. The sensor nodes of the server rack constitute the vertices of the graph. The node features are the vibration energy distribution features and typical spectrum pattern features. The graph edges are defined as physical adjacency relationships. Through multi-layer convolution and attention mechanisms, the spatiotemporal correlation is learned in real time, and the seismic event confidence score Cev is output to classify real earthquake events and equipment self-vibration.
7. The seismic monitoring method for a server rack according to claim 1, characterized in that: Based on the aforementioned vibration energy distribution characteristics, typical spectral pattern characteristics, and seismic event confidence Cev, historical baseline energy Eref and spectral template constant Fref[i] are extracted. α, β, and γ are set as weighting coefficients, and the comprehensive seismic risk index SRI is calculated accordingly.
8. The seismic monitoring method for a server rack according to claim 1, characterized in that: A pre-set seismic risk threshold Tth is established. When the comprehensive seismic risk index SRI is greater than or equal to the seismic risk threshold Tth, the system issues a high-risk warning command and automatically executes the following emergency response: Trigger the enhanced mode of the cabinet vibration isolation support, push an alarm to the upper-level monitoring platform, and initiate the UPS safety shutdown procedure; When the comprehensive seismic risk index SRI is less than the seismic risk threshold Tth, the routine monitoring status is maintained.
9. The seismic monitoring method for a server rack according to claim 1, characterized in that: The system records key parameters Nsc, vibration energy distribution characteristics, typical spectral pattern characteristics, earthquake event confidence level Cev, and comprehensive seismic risk index SRI in real time during the monitoring process. The monitoring logs and model outputs are asynchronously uploaded to the cloud database through the edge gateway, supporting subsequent big data analysis and online model retraining.
10. The seismic monitoring method for a server rack according to claim 1, characterized in that: Based on cloud-based big data analysis results, the monitoring plan is dynamically adjusted, including: optimizing sensor deployment density and location, fine-tuning ICA and Wiener filter parameters online, regularly updating graph neural network model weights, enhancing the performance of rack vibration isolation devices, and regularly organizing seismic emergency drills for maintenance personnel to continuously improve the seismic monitoring capabilities of data center server racks.