An AI server cooling water pump vibration amount adaptive calibration test method and system

Abstract

The application relates to an AI server cooling water pump vibration quantity self-adaptive calibration test method and system, an intelligent test platform integrates a six-dimensional force sensor and a temperature compensation exciter, the rigidity of an automatic calibration platform is calibrated, a water pump-pipeline system transfer function is acquired, a three-dimensional mathematical model of rotating speed-lift-moded frequency is established, a three-axis MEMS accelerometer array is arranged at sensitive parts such as a water pump bearing seat and a motor shell, vibration, current and pressure signals are synchronously collected, a variational mode decomposition is combined with an adaptive S transform to perform time-frequency domain signal processing, 128-dimensional full-frequency domain characteristic vectors containing modulation side frequency bands are extracted, vibration acceleration, speed and displacement effective values are calculated through a frequency domain integral algorithm, and current harmonic interference is corrected by using an electromagnetic vibration compensation model, finally, a vibration health degree evaluation model containing 18 characteristic parameters is established based on a support vector machine, the test system comprises an intelligent test platform unit, a multi-source sensing unit, an edge computing unit and a data management unit, microsecond-level synchronous collection and GB-level data throughput are realized through a time-sensitive network, online incremental learning and automatic generation of a test report are supported, and the application has the effect of improving water pump vibration quantity test precision.

Description

Technical Field

[0001] This application relates to the field of water pump testing, and in particular to an adaptive calibration test method and system for vibration of a water pump used for cooling AI servers. Background Technology

[0002] With the rapid development of artificial intelligence technology, the computing scale of AI server clusters is growing exponentially, and the high heat generated during their operation places stringent demands on cooling systems. As the core power component of the AI ​​server liquid cooling system, the stable operation of the water pump directly determines cooling efficiency and server reliability. Water pump vibration is not only a key indicator of its operating status, but abnormal vibration can also lead to seal failures, bearing wear, and other malfunctions, resulting in coolant leaks or even server downtime. Therefore, accurate testing of water pump vibration to achieve early fault warning has become a necessary step in ensuring the stable operation of AI data centers.

[0003] Currently, traditional water pump vibration testing methods have significant limitations. Regarding test environment construction, existing test platforms mostly employ fixed-stiffness structures, which cannot adapt to the dynamic characteristics of different water pump models, leading to boundary condition errors affecting test results. In the signal acquisition stage, single-point or low-precision sensor deployments cannot comprehensively capture the coupled vibration characteristics of multiple parts of the water pump, and lack synchronous monitoring of multiple physical quantities such as current and pressure, making it impossible to isolate interference factors such as electromagnetic vibration. At the data processing level, conventional Fourier transforms are insufficient for analyzing complex modulation signals, and simple threshold judgments cannot meet the high-precision diagnostic requirements of AI-based water pumps, resulting in large errors in vibration calculation and low fault diagnosis accuracy.

[0004] To address the aforementioned issues, existing technologies struggle to accurately test and assess the vibration levels of AI server cooling water pumps. There is an urgent need for a testing method and system with adaptive calibration capabilities and multi-source data integration. This system should leverage intelligent sensing and advanced algorithms to effectively improve the accuracy of vibration testing and the reliability of fault diagnosis, filling a technological gap in the testing of dedicated cooling equipment for AI servers. Summary of the Invention

[0005] To improve the accuracy of water pump vibration measurement, this application provides an adaptive calibration test method and system for the vibration of an AI server cooling water pump.

[0006] The adaptive calibration test method and system for vibration of an AI server cooling water pump provided in this application adopts the following technical solution:

[0007] First aspect

[0008] A method for testing the vibration of an AI server cooling water pump includes the following steps:

[0009] S1. Adaptive calibration start: The water pump under test is installed on an intelligent test platform integrating a six-dimensional force sensor. The platform stiffness matrix is ​​automatically calibrated through a closed-loop feedback control system. The initial modal parameters of the water pump under no-load conditions are collected simultaneously, and a three-dimensional mathematical model containing the mapping relationship between speed, head and modal frequency is established.

[0010] S2. Synchronous acquisition of multimodal signals: A triaxial MEMS accelerometer array is arranged at the radial and axial measuring points of the water pump bearing housing, the motor rotor housing and the impeller volute according to the ISO 1940 balance level requirements. A synchronous trigger data acquisition card is used to acquire the vibration time domain signal at a sampling rate of not less than 20kHz. At the same time, the water pump drive current waveform and inlet and outlet pressure real-time data are acquired simultaneously.

[0011] S3. Time-frequency domain joint signal processing: The original vibration signal is subjected to variational mode decomposition. The decomposition parameters are optimized in the range of mode 2-8 by adaptive particle swarm optimization algorithm. The intrinsic mode functions with kurtosis values ​​greater than 2.5 are selected by combining the kurtosis criterion. The time-frequency energy distribution spectrum is constructed based on the adaptive S-transform. The three-dimensional mathematical model established in step S1 is used to identify the fundamental frequency vibration component and extract the full-frequency domain feature vector containing the modulation sideband.

[0012] S4. Multi-physical quantity fusion calculation: Based on the accelerometer sensitivity matrix, the time-domain vibration signal is transformed into coordinates. The effective values ​​of vibration acceleration, vibration velocity, and vibration displacement are calculated simultaneously using the frequency domain integration algorithm. An electromagnetic vibration compensation model is constructed by establishing the cross-correlation function between the driving current spectrum and the vibration signal. The compensation coefficient matrix is ​​generated by least squares fitting of the current harmonic components with a correlation coefficient greater than 0.3, and the calculation results are corrected.

[0013] By adopting the above technical solutions, an intelligent testing platform integrating six-dimensional force sensors is used to achieve adaptive calibration startup. A three-dimensional mathematical model of pump speed-head-modal frequency is established, solving the problem of large boundary condition errors in traditional testing platforms. This allows for precise matching of the dynamic characteristics of different pumps. Simultaneous acquisition of multi-modal signals combined with monitoring of multiple physical quantities comprehensively captures pump vibration characteristics and isolates interference factors. Time-frequency domain joint signal processing and multi-physical quantity fusion calculation effectively analyze complex vibration signals, improve the accuracy of vibration quantity calculation, and provide a reliable data foundation for pump health diagnosis.

[0014] Preferably, the intelligent testing platform is equipped with a temperature-compensated vibrator, which acquires the transfer function matrix of the water pump-pipeline system through 10-1000Hz sweep frequency excitation during the startup phase, and is used to correct the boundary conditions of the subsequently acquired vibration signals.

[0015] By adopting the above technical solution, the temperature-compensated vibrator configured in the intelligent testing platform obtains the transfer function matrix of the water pump-pipeline system through 10-1000Hz sweep frequency excitation. It can correct the boundary conditions of the vibration signal according to the changes in ambient temperature and the dynamic characteristics of the pipeline, eliminate the influence of external factors on the test results, and further improve the accuracy and reliability of the test.

[0016] Preferably, the installation position of the accelerometer array is precisely calibrated in three dimensions using a laser rangefinder to form a sensor layout matrix containing position information. The installation angle error of each sensor is controlled within ±0.5°, meeting the ISO 20816 vibration measurement position accuracy requirements.

[0017] By adopting the above technical solution, the installation position of the accelerometer array is accurately calibrated using a laser rangefinder, and the sensor installation angle error is controlled within ±0.5°, which meets the high-precision measurement requirements of international standards, ensures the accuracy of the spatial position of the collected vibration signal, avoids signal distortion caused by sensor installation deviation, and improves the effectiveness of vibration data acquisition.

[0018] Preferably, when extracting the full-frequency domain feature vector, a ±3 times frequency shift analysis bandwidth centered on the fundamental frequency is defined, and an improved Teager energy operator is used to calculate the instantaneous energy of each frequency component, thereby constructing a 128-dimensional vibration state vector containing time-domain, frequency-domain, and time-frequency-domain features.

[0019] By adopting the above technical solution, when extracting feature vectors in the full frequency domain, the analysis bandwidth is defined with the fundamental frequency as the center, and the instantaneous energy is calculated using the improved Teager energy operator to construct a 128-dimensional vibration state vector. This can completely extract the time domain, frequency domain, and time-frequency domain features of the vibration signal, effectively identify early weak fault features, and improve the sensitivity and accuracy of fault diagnosis.

[0020] Preferably, the vibration calculation results include:

[0021] Temporal characteristic quantities: peak value, effective value, kurtosis value, margin factor;

[0022] Frequency domain characteristic quantities: fundamental frequency amplitude ratio, sideband energy ratio, octave band energy distribution;

[0023] Time-frequency domain characteristic quantities: energy centroid frequency, root mean square frequency, frequency variance.

[0024] By adopting the above technical solution, multiple time-domain, frequency-domain, and time-frequency-domain characteristic quantities are defined as vibration calculation results, which comprehensively characterize the vibration state of the water pump from multiple dimensions. Compared with single index analysis, it can reflect the operating status of the water pump more meticulously and accurately, and provide rich parameter basis for subsequent fault diagnosis and health assessment.

[0025] Preferably, a vibration threshold self-learning module is set in the data processing stage to train a support vector machine classifier through historical test data and establish a water pump vibration health assessment model containing 18 feature parameters. The feature parameters include all the above-mentioned preferred feature quantities and water pump operating condition parameters.

[0026] By adopting the above technical solution, by setting a vibration threshold self-learning module, and using historical test data to train an SVM classifier, a water pump vibration health assessment model containing multiple feature parameters is established, realizing the autonomous learning and intelligent diagnosis of the test system. The diagnostic threshold can be dynamically adjusted according to different water pump characteristics, thereby improving the adaptability and accuracy of fault diagnosis.

[0027] Second aspect

[0028] A system for testing the vibration of cooling water pumps used in the aforementioned AI server includes:

[0029] a. Intelligent testing platform unit: integrates a six-dimensional force sensor array, a temperature-compensated vibrator, and a laser calibration module, which can construct a dynamic coordinate system for water pump vibration testing and automatically complete platform stiffness calibration and transfer function matrix measurement;

[0030] b. Multi-source sensing unit: includes a 12-channel triaxial MEMS accelerometer, a high-frequency current transformer and a pressure transmitter, and achieves microsecond-level time alignment through a synchronous clock module. The sampling rate is not less than 20kHz, which meets the ISO 13373 vibration signal acquisition standard.

[0031] c. Edge computing unit: Equipped with an FPGA real-time signal processing module, it has a built-in VMD noise reduction algorithm module, an adaptive S-transform engine, and a multi-physical quantity fusion computing unit, supporting real-time time-frequency domain analysis and characteristic parameter calculation of vibration signals;

[0032] d. Data Management Unit: Based on a cloud computing platform, a vibration characteristic database is built, integrating a health assessment model training module and an automatic test report generation system, which can realize the storage of test data, model training, and visualization report output.

[0033] By adopting the above technical solutions, and through the collaborative work of the intelligent testing platform unit, multi-source sensing unit, edge computing unit, and data management unit, a complete water pump vibration testing system has been constructed. A closed loop is formed from hardware sensing to data processing, storage, and analysis, realizing the automation and intelligence of the testing process and improving testing efficiency and data management capabilities.

[0034] Preferably, the edge computing unit and the data management unit use time-sensitive network communication, with signal transmission delay controlled within 50 microseconds and data processing throughput not less than 1GB / s, to ensure the real-time performance and integrity of the test data.

[0035] By adopting the above technical solution and using Time-Sensitive Network (TSN) communication, the signal transmission delay between the edge computing unit and the data management unit is controlled within 50 microseconds, ensuring that the data processing throughput is not less than 1GB / s, meeting the real-time transmission and processing requirements of a large amount of vibration data, ensuring the timeliness and integrity of test data, and providing a guarantee for real-time monitoring and rapid diagnosis.

[0036] Preferably, the sensitivity matrix of the triaxial MEMS accelerometer array is pre-calibrated using a national standard vibration calibration device, with the sensitivity error controlled within ±1.5% and the frequency response range covering 0.1Hz-10kHz.

[0037] By adopting the above technical solution, the triaxial MEMS accelerometer array is subjected to national standard vibration calibration, which controls the sensitivity error within ±1.5% and covers the frequency response range of 0.1Hz-10kHz. This ensures the high accuracy and wide frequency applicability of the sensor data, improves the accuracy and reliability of vibration signal acquisition, and lays a solid foundation for subsequent data processing and analysis.

[0038] Preferably, the vibration health assessment model training module supports online incremental learning and can dynamically update the compensation coefficient matrix and SVM classifier parameters based on real-time test data, thereby realizing the self-calibration and flexibility of the test system.

[0039] By adopting the above technical solution, the vibration health assessment model training module supports online incremental learning and can dynamically update the compensation coefficient matrix and SVM classifier parameters according to real-time test data. This enables the test system to have self-calibration and self-optimization capabilities, adapt to changes in pump operating status and test environment, and continuously maintain high diagnostic accuracy and reliability.

[0040] In summary, this application includes at least one of the following beneficial technical effects:

[0041] 1. By integrating a six-dimensional force sensor into an intelligent testing platform and a temperature-compensated vibrator, the platform stiffness matrix is ​​automatically calibrated and the transfer function matrix of the pump-pipeline system is obtained. This solves the problem that traditional fixed stiffness platforms cannot adapt to the dynamic characteristics of different pumps, reducing boundary condition errors by more than 60%. The acquisition of no-load modal parameters and the establishment of a three-dimensional mathematical model achieve precise mapping of speed, head, and modal frequency, providing a reliable dynamic coordinate system for subsequent vibration signal analysis and significantly improving the versatility and consistency of testing different pump models.

[0042] 2. Variational Mode Decomposition (VMD) and Adaptive Particle Swarm Optimization (PSO) are employed to optimize decomposition parameters. Valid modes are selected using the kurtosis criterion, suppressing noise interference while preserving early fault characteristics. A time-frequency energy map constructed based on the adaptive S-transform enables a visual representation of the vibration signal in a three-dimensional space of time, frequency, and energy. Combined with an improved Teager energy operator, a 128-dimensional full-frequency domain feature vector is extracted. Compared to the traditional Fourier transform, the sensitivity for identifying weak fault characteristics such as modulation sidebands is improved by 40%, effectively solving the problem of analyzing complex modulation signals.

[0043] 3. The edge computing unit and data management unit communicate using Time-Sensitive Networking (TSN), with signal transmission latency controlled within 50 microseconds and data throughput ≥1GB / s, ensuring real-time processing and characteristic parameter calculation of multi-channel vibration signals at a 20kHz sampling rate (response time ≤200 microseconds). Combined with the cloud computing platform's database management and visual report generation system, the entire process from signal acquisition and real-time analysis to health assessment is automated, significantly improving testing efficiency and data management capabilities. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the execution steps of an adaptive calibration test method for vibration of an AI server cooling water pump according to an embodiment of this application.

[0045] Figure 2 This is a block diagram of the time-frequency domain joint signal processing flow in an adaptive calibration test method for vibration of an AI server cooling water pump according to an embodiment of this application.

[0046] Figure 3 This is a flowchart of the multi-physical quantity fusion calculation process in an adaptive calibration test method for vibration of an AI server cooling water pump according to an embodiment of this application.

[0047] Figure 4 This is a block diagram illustrating the overall testing principle of an adaptive calibration test method and system for vibration of an AI server cooling water pump, as described in an embodiment of this application. Detailed Implementation

[0048] The following is in conjunction with the appendix Figure 1-4 This application will be described in further detail.

[0049] This application discloses an adaptive calibration test method and system for the vibration of an AI server cooling water pump.

[0050] Reference Figure 1 A method for testing the vibration of an AI server cooling water pump includes the following steps:

[0051] S1. Adaptive Calibration Start-up: The water pump under test is installed on an intelligent test platform integrating a six-dimensional force sensor. The intelligent test platform is equipped with a temperature-compensated vibrator. During the start-up phase, the transfer function matrix of the water pump-pipeline system is obtained through 10-1000Hz sweep frequency excitation. This matrix is ​​used to correct the boundary conditions of the subsequently acquired vibration signals. The platform stiffness matrix is ​​automatically calibrated through a closed-loop feedback control system. The initial modal parameters of the water pump under no-load conditions are acquired simultaneously, and a three-dimensional mathematical model containing the mapping relationship between speed, head, and modal frequency is established.

[0052] Specifically, the collected speed, head, and modal frequency data are imported into the Python Scikit-learn library, and a multiple linear regression algorithm is used for fitting to establish a three-dimensional mathematical model that includes the mapping relationship between speed, head, and modal frequency. The model parameters are optimized through cross-validation to ensure the model's coefficient of determination R0. 2 A value greater than 0.98 ensures the model's accuracy and generalization ability. In actual testing, based on the real-time speed and head of the water pump, this model can quickly predict the corresponding modal frequencies, providing a reference for subsequent vibration analysis.

[0053] S2. Synchronous acquisition of multimodal signals: Refer to Figure 2 To comprehensively monitor the vibration characteristics of the water pump at different locations, a triaxial MEMS accelerometer array is arranged at radial and axial measuring points on the pump bearing housing, the motor rotor housing, and the impeller volute. For example, one accelerometer is placed radially and axially on the bearing housing, two symmetrically on the motor rotor housing, and two near the blades on the impeller volute. This arrangement conforms to ISO 1940 balance requirements. A synchronous trigger data acquisition card is used to acquire the vibration time-domain signal at a sampling rate of at least 20kHz. Simultaneously, the pump drive current waveform and real-time inlet and outlet pressure data are acquired. The installation position of the accelerometer array is precisely calibrated in three dimensions using a laser rangefinder, forming a sensor layout matrix containing position information. The installation angle error of each sensor is controlled within ±0.5°, meeting the ISO 20816 vibration measurement position accuracy requirements. For the data acquisition card, a synchronous trigger data acquisition card, such as the NIPXIe-4499, can be used. This card has multi-channel synchronous acquisition capabilities and a maximum sampling rate of 102.4kS / s, meeting the requirement of a sampling rate of at least 20kHz. Configure 16-24 analog input channels, of which 12-18 channels are used to connect to accelerometers (each triaxial accelerometer occupies 3 channels), 2-4 channels are used to acquire current signals, and 2 channels are used to acquire pressure signals, ensuring that all types of signals can be simultaneously accessed by the acquisition system.

[0054] Furthermore, setting the sampling rate of the data acquisition card to 24kHz ensures sufficient signal detail is captured while avoiding data redundancy and processing pressure caused by excessively high sampling rates. The acquisition card's trigger mode is set to external synchronous trigger. By connecting a high-precision synchronization clock module (such as a GPS synchronization clock or a dedicated 10MHz synchronization clock), microsecond-level time alignment of all channel signals is achieved, ensuring consistency of multimodal signals in the time dimension.

[0055] Simultaneously, the entire system needs to be calibrated. After all sensors are installed, the entire multimodal signal acquisition system is jointly calibrated. By simulating water pump operation, standard vibration, current, and pressure signals are input to check whether the output of each channel of the acquisition system is consistent with the input signals. Calibration software is used to adjust parameters such as gain and offset of the acquisition card to ensure that the overall measurement accuracy of the system meets the requirements. At the same time, the synchronization effect of the synchronization clock module is verified. By comparing the time difference of the same trigger signal acquired from different channels, synchronization parameters are adjusted to ensure accurate synchronous acquisition of multimodal signals.

[0056] S3. Joint Time-Frequency Domain Signal Processing: Variational mode decomposition is performed on the original vibration signal. The decomposition parameters are optimized within the range of modes 2-8 using an adaptive particle swarm optimization algorithm. Eigenmode functions with kurtosis values ​​greater than 2.5 are selected based on the kurtosis criterion. A time-frequency energy distribution map is constructed based on the adaptive S-transform. The three-dimensional mathematical model established in step S1 is used to identify the fundamental frequency vibration component and extract the full-frequency domain feature vector containing the modulation sideband. When extracting the full-frequency domain feature vector, a ±3 times re-frequency analysis bandwidth centered on the fundamental frequency is defined. The instantaneous energy of each frequency component is calculated using the improved Teager energy operator, and a 128-dimensional vibration state vector containing time-domain, frequency-domain, and time-frequency domain features is constructed.

[0057] Specifically, the acquired raw vibration signals often contain noise and interference. First, they are processed to remove the DC component and slow changing trends, making the signal more focused on the vibration characteristics. Then, a bandpass filter is used, and combined with the frequency characteristics of the water pump vibration, the filter passband is set to 0.1Hz-10kHz to suppress low-frequency environmental noise and high-frequency electromagnetic interference, while retaining the effective vibration frequency band signal.

[0058] Furthermore, variational mode decomposition (VMD) is the core step in decomposing the original signal into multiple intrinsic mode functions (IMFs). VMD adaptively decomposes complex signals into a finite number of bandwidth-constrained modal components by constructing and solving variational models. Its core variational problem is expressed as:

[0059]

[0060] Among them, u k To decompose the modal function set, wk Let be the center frequency of each mode, and s(t) be the original vibration signal.

[0061] To achieve optimal decomposition results, an adaptive particle swarm optimization (PSO) algorithm is used to optimize the number of decomposition modes K in VMD (set within the range of 2-8). The PSO algorithm simulates the foraging behavior of a flock of birds, searching for the optimal solution in the solution space. In each iteration, the particles are adjusted based on their historical best position p. best and the global optimal position g best Update position, update formula as follows:

[0062] v id (t+1)=ω·v id (t)+c1r 1d (t)[p id (t)-x i x id [(t+1)]=x id (t+1)=x id (t)+v id (t+1)

[0063] Among them, v id Let x be the particle velocity. id Let ω be the particle position, ω be the inertia weight, c1 be the learning factor, and r be the particle position. 1d It is a random number.

[0064] Using the kurtosis value of each IMF after decomposition as the optimization objective, the kurtosis calculation formula is as follows:

[0065]

[0066] Where, x i For signal sample values, σ is the mean and σ is the standard deviation. IMFs with kurtosis values ​​greater than 2.5 are selected, and modal components containing fault characteristics are retained for subsequent analysis.

[0067] Furthermore, the signal analysis method based on adaptive S-transform and time-frequency localization characteristics is defined as follows:

[0068]

[0069] To adapt to the characteristics of water pump vibration signals, an adaptive S-transform is adopted. By optimizing the scaling parameters of the S-transform, the time-frequency resolution for different frequency components is enhanced. The selected IMF is subjected to an adaptive S-transform to obtain a time-frequency matrix. With time as the horizontal axis, frequency as the vertical axis, and amplitude as the color depth, a time-frequency energy distribution map is constructed to intuitively display the energy distribution characteristics of the vibration signal in the time-frequency domain.

[0070] Correspondingly, using the three-dimensional mathematical model established in step S1 and combined with the time-frequency energy spectrum, the fundamental frequency vibration component of the vibration signal is identified. A ±3 times rotational frequency analysis bandwidth centered on the fundamental frequency is defined, and within this range, the instantaneous energy of each frequency component is calculated using the improved Teager energy operator. The formula for the improved Teager energy operator is:

[0071] Ψ[x(n)]=x 2 (n)-x(n-1)x(n+1)

[0072] By calculating instantaneous energy, a 128-dimensional vibration state vector containing time-domain (such as peak value and RMS value), frequency-domain (such as fundamental frequency amplitude and sideband energy), and time-frequency domain (such as energy centroid frequency) features is extracted to comprehensively characterize the pump vibration characteristics and provide rich feature information for fault diagnosis.

[0073] S4. Multi-physical quantity fusion calculation: Refer to Figure 3 The time-domain vibration signal is transformed into coordinates based on the accelerometer sensitivity matrix. The effective values ​​of vibration acceleration, vibration velocity, and vibration displacement are calculated simultaneously using a frequency domain integration algorithm. An electromagnetic vibration compensation model is constructed by establishing the cross-correlation function between the driving current spectrum and the vibration signal. The compensation coefficient matrix is ​​generated by least squares fitting of the current harmonic components with a correlation coefficient greater than 0.3, and the calculation results are corrected.

[0074] Based on the accelerometer sensitivity matrix, the voltage signal output by the sensor is converted into an actual physical quantity (acceleration). The specific steps are as follows:

[0075] S401. Sensitivity Matrix Calibration

[0076] The sensitivity matrix for each triaxial MEMS accelerometer is as follows: Among them, s x s y s z Sensitivity for each of the three axes (unit: m·s) -2 The system is calibrated using a national standard vibration calibration device (such as the vibration calibration system of the National Institute of Metrology, China) to ensure that the sensitivity error is controlled within ±1.5% and the frequency response range covers 0.1Hz-10kHz.

[0077] Furthermore, the coordinate transformation formula is as follows: Let the original voltage signal acquired by the sensor be V(t) = [V x (t),V y (t),V z (t)] T Then the actual acceleration signal is: a(t)=S·V(t).

[0078] Among them, a(t)=[ax (t),a y (t),a z (t)] T This is a time-domain signal of triaxial acceleration.

[0079] S402. For the frequency domain integral factor, where,

[0080] Vibration velocity frequency domain signal: (First integration, where j is the imaginary unit);

[0081] Vibration displacement frequency domain signal: (Double integral)

[0082] Subsequently, through inverse Fourier transform and time-domain RMS value calculation, inverse Fourier transform is performed on V(f) and D(f) to obtain time-domain signals v(t) and d(t), and the RMS value is calculated:

[0083] RMS acceleration value:

[0084] RMS speed:

[0085] Effective value of displacement:

[0086] Where T is the signal acquisition time.

[0087] S403. Furthermore, the construction of the electromagnetic vibration compensation model must be carried out through the following steps;

[0088] S4031. Current Signal Spectrum Analysis: Perform FFT on the driving current signal i(t) to obtain the current frequency I(f), and extract the harmonic components If. k (f) = I(kf0), where f0 is the fundamental frequency (the electrical frequency corresponding to the pump speed).

[0089] S4032. Cross-correlation coefficient calculation: Calculate the cross-correlation coefficient between current harmonic components and vibration signals (taking acceleration as an example):

[0090] Among them, i k (t) is the time-domain signal of the k-th harmonic, Cov is the covariance, and σ a σ ik For the standard deviation, select |ρ k The harmonic component with a value of 0.3 is the main source of interference.

[0091] S4033. Least Squares Fitting Compensation Parameters

[0092] Let the compensation coefficient matrix be: K = [k 1, k 2,…,k m ], m is the number of effective interference harmonics, and a linear compensation model is established: a 补偿 (t)=a(t)-K·i 干扰 (t)

[0093] The optimal K is found by minimizing the kurtosis (or mean square error) of the compensated signal.

[0094]

[0095] S4034. For the frequency domain integral filter settings, a 50-stage Butterworth low-pass filter is used, with a cutoff frequency of [missing information]. (f s =24kHz is the sampling rate), to suppress high-frequency noise introduced during the integration process.

[0096] S4035. Compensation effect verification: Taking a certain type of water pump as an example: The effective value of vibration acceleration without compensation is 1.2 m / s². -2 The cross-correlation coefficient of the 5th harmonic (f = 250 Hz) was detected to be 0.42, and after compensation, the effective value decreased to 0.9 m / s. -2 The kurtosis value decreased from 3.2 to 2.1, and surface electromagnetic interference was effectively suppressed.

[0097] Based on the above process, the vibration calculation results include:

[0098] Temporal characteristic quantities: peak value, effective value, kurtosis value, margin factor;

[0099] Frequency domain characteristic quantities: fundamental frequency amplitude ratio, sideband energy ratio, octave band energy distribution;

[0100] Time-frequency domain characteristic quantities: energy centroid frequency, root mean square frequency, frequency variance.

[0101] In the data processing stage, a vibration threshold self-learning module is set up. A support vector machine classifier is trained using historical test data to establish a water pump vibration health assessment model containing 18 feature parameters. The feature parameters include time-domain features, frequency-domain features, and all features mentioned above, as well as water pump operating condition parameters. The vibration health assessment model training module supports online incremental learning and can dynamically update the compensation coefficient matrix and SVM classifier parameters according to real-time test data, realizing the self-calibration and flexibility of the test system.

[0102] Furthermore, refer to Figure 4 A test system for testing the vibration of the cooling water pump for the aforementioned AI server includes the following modules:

[0103] a. Intelligent testing platform unit: integrates a six-dimensional force sensor array, a temperature-compensated vibrator, and a laser calibration module, which can construct a dynamic coordinate system for water pump vibration testing and automatically complete platform stiffness calibration and transfer function matrix measurement;

[0104] b. Multi-source sensing unit: includes a 12-channel triaxial MEMS accelerometer, a high-frequency current transformer and a pressure transmitter, and achieves microsecond-level time alignment through a synchronous clock module. The sampling rate is not less than 20kHz, which meets the ISO 13373 vibration signal acquisition standard.

[0105] c. Edge Computing Unit: Equipped with an FPGA real-time signal processing module, it incorporates a VMD noise reduction algorithm module, an adaptive S-transform engine, and a multi-physical quantity fusion calculation unit. Based on the sensor layout matrix and sensitivity matrix, it converts the vibration signals of each measuring point into acceleration, velocity, and displacement signals in the global coordinate system. It calculates the effective values ​​using the frequency domain integration method (accuracy ≤ 0.5%) and corrects current harmonic interference through an electromagnetic vibration compensation model (response time ≤ 200 microseconds). It supports real-time time-frequency domain analysis and characteristic parameter calculation of vibration signals. The edge computing unit and the data management unit use time-sensitive network communication, with signal transmission delay controlled within 50 microseconds and data processing throughput not less than 1GB / s, ensuring the real-time performance and integrity of the test data.

[0106] d. Data Management Unit: Based on a cloud computing platform, a vibration characteristic database is built, integrating a health assessment model training module and an automatic test report generation system. It can realize the storage of test data, model training, and visualization report output. The report format supports PDF / A-3 (long-term archiving) and HTML5 (web-based interactive), and includes electronic signatures and timestamps (compliant with ISO17025 standard). The implementation principle of the adaptive calibration test method and system for vibration of an AI server cooling water pump in this application embodiment is as follows: The intelligent test platform unit uses a six-dimensional force sensor, a temperature-compensated exciter, and a laser calibration module to construct a dynamic coordinate system and automatically calibrate the platform stiffness, obtaining the transfer function matrix of the water pump-pipeline system to provide accurate mechanical boundary conditions for testing; the multi-source sensing unit uses sensors such as a three-axis MEMS accelerometer to reasonably arrange sensors at key parts of the water pump, combined with the IEEE 1588 protocol to realize the synchronous acquisition and time alignment of multi-modal signals, comprehensively acquiring data such as vibration, current, and pressure; the edge computing unit is based on FPGA hardware architecture and uses algorithms such as VMD and adaptive S-transform to preprocess, extract features, and perform multi-physical quantity fusion calculation on the acquired signals, quickly and accurately generating results such as a 128-dimensional vibration state vector; the data management unit uses storage technologies such as InfluxDB to establish a vibration feature database, uses TensorFlow to train an SVM classifier to realize offline training and online incremental learning of the health assessment model, and automatically generates a test report that meets the standards. Each unit works collaboratively through communication technologies such as time-sensitive networking, and data is transmitted and optimized in an orderly manner between units, ultimately achieving high-precision and intelligent testing and health assessment of the vibration of the AI ​​server cooling water pump.

[0107] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An AI server cooling water pump vibration amount test method, characterized by, Includes the following steps: S1. Adaptive calibration start: The water pump under test is installed on an intelligent test platform with integrated six-dimensional force sensors. The platform stiffness matrix is ​​automatically calibrated through a closed-loop feedback control system. The initial modal parameters of the water pump under no-load conditions are collected simultaneously, and a three-dimensional mathematical model containing the mapping relationship between speed, head and modal frequency is established. S2. Synchronous acquisition of multimodal signals: A triaxial MEMS accelerometer array is arranged at the radial and axial measuring points of the water pump bearing housing, the motor rotor housing and the impeller volute according to the ISO 1940 balance level requirements. A synchronous trigger data acquisition card is used to acquire the vibration time domain signal at a sampling rate of not less than 20kHz. At the same time, the water pump drive current waveform and inlet and outlet pressure real-time data are acquired simultaneously. S3. Time-frequency domain joint signal processing: The original vibration signal is subjected to variational mode decomposition. The decomposition parameters are optimized in the range of mode 2-8 by adaptive particle swarm optimization algorithm. The intrinsic mode functions with kurtosis values ​​greater than 2.5 are selected by combining the kurtosis criterion. The time-frequency energy distribution spectrum is constructed based on the adaptive S-transform. The three-dimensional mathematical model established in step S1 is used to identify the fundamental frequency vibration component and extract the full-frequency domain feature vector containing the modulation sideband. S4. Multi-physical quantity fusion calculation: Based on the accelerometer sensitivity matrix, the time-domain vibration signal is transformed into coordinates. The effective values ​​of vibration acceleration, vibration velocity, and vibration displacement are calculated simultaneously using the frequency domain integration algorithm. An electromagnetic vibration compensation model is constructed by establishing the cross-correlation function between the driving current spectrum and the vibration signal. The compensation coefficient matrix is ​​generated by least squares fitting of the current harmonic components with a correlation coefficient greater than 0.3, and the calculation results are corrected.

2. The test method of claim 1, wherein: The intelligent testing platform is equipped with a temperature-compensated vibrator. During the startup phase, it acquires the transfer function matrix of the water pump-pipeline system through 10-1000Hz sweep frequency excitation, which is used to correct the boundary conditions of the subsequently acquired vibration signals.

3. The test method of claim 1, wherein: The installation position of the accelerometer array is precisely calibrated in three dimensions using a laser rangefinder, forming a sensor layout matrix containing position information. The installation angle error of each sensor is controlled within ±0.5°, meeting the ISO 20816 vibration measurement position accuracy requirements.

4. The test method of claim 1, wherein: When extracting the full-frequency domain feature vector, a ±3 times frequency shift analysis bandwidth centered on the fundamental frequency is defined. An improved Teager energy operator is used to calculate the instantaneous energy of each frequency component, and a 128-dimensional vibration state vector containing time-domain, frequency-domain, and time-frequency-domain features is constructed.

5. The test method of claim 1, wherein, The vibration calculation results include: Temporal characteristic quantities: peak value, effective value, kurtosis value, margin factor; Frequency domain characteristic quantities: fundamental frequency amplitude ratio, sideband energy ratio, octave band energy distribution; Time-frequency domain characteristic quantities: energy centroid frequency, root mean square frequency, frequency variance.

6. The test method according to any one of claims 1 to 5, characterized in that: In the data processing stage, a vibration threshold self-learning module is set up, and a support vector machine classifier is trained through historical test data to establish a water pump vibration health assessment model containing 18 feature parameters.

7. The AI ​​server cooling water pump vibration testing system according to any one of claims 1-6, characterized in that, include: a. Intelligent testing platform unit: integrates a six-dimensional force sensor array, a temperature-compensated exciter, and an exciter... The cursor calibration module can construct a dynamic coordinate system for pump vibration testing and automatically complete platform stiffness calibration and transfer function matrix measurement. b. Multi-source sensing unit: includes a 12-channel triaxial MEMS accelerometer, a high-frequency current transformer and a pressure transmitter, and achieves microsecond-level time alignment through a synchronous clock module. The sampling rate is not less than 20kHz, which meets the ISO 13373 vibration signal acquisition standard. c. Edge computing unit: Equipped with an FPGA real-time signal processing module, it has a built-in VMD noise reduction algorithm module, an adaptive S-transform engine, and a multi-physical quantity fusion computing unit, supporting real-time time-frequency domain analysis and characteristic parameter calculation of vibration signals; d. Data Management Unit: Based on a cloud computing platform, a vibration characteristic database is built, integrating a health assessment model training module and an automatic test report generation system, which can realize the storage of test data, model training, and visualization report output.

8. The test system of claim 7, wherein: The edge computing unit and the data management unit use time-sensitive network communication, with signal transmission delay controlled within 50 microseconds and data processing throughput not less than 1GB / s, to ensure the real-time performance and integrity of the test data.

9. The testing system according to claim 7, characterized in that: The sensitivity matrix of the triaxial MEMS accelerometer array is pre-calibrated using a vibration calibration device, with the sensitivity error controlled within ±1.5%, and the frequency response range covering 0.1Hz-10kHz.

10. The test system of claim 7, wherein: The vibration health assessment model training module supports online incremental learning and can dynamically update the compensation coefficient matrix and SVM classifier parameters based on real-time test data, thereby enabling the self-calibration and customization of the test system.

Images