A method and system for diagnosing power supply failures in AI servers

By performing multi-dimensional analysis of the steady-state voltage value of AI server power supplies and quantifying the aging index of capacitors and connectors, the problem of the difficulty in dissecting the combined voltage drop trend caused by capacitor aging and connector degradation is solved. This enables accurate diagnosis and early warning of hidden power supply faults, improving operation and maintenance efficiency and system stability.

CN123114831APending Publication Date: 2026-11-17GUIZHOU WEIQING DEV GRP CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611422212.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In scenarios where AI servers operate under long-term high loads, the combined voltage drop caused by aging capacitors and degraded connectors inside the power supply is difficult to resolve. Existing technologies cannot accurately distinguish the source of the fault, leading to ambiguity in the diagnosis of hidden faults.

Method used

By acquiring steady-state voltage values ​​over multiple natural days, and using weighted least squares fitting and residual sequence analysis, the aging characteristics of capacitors and connectors are decoupled, the aging index of capacitors and connectors is quantified, and dynamic threshold determination is performed in conjunction with the ambient temperature rise gradient to achieve accurate fault diagnosis.

Benefits of technology

Successfully decoupled capacitor aging from connector contact problems, enabling accurate identification and early warning of hidden power supply faults, thus improving the efficiency of data center operation and maintenance and the operational stability of server systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114831A_ABST
    Figure CN123114831A_ABST
Patent Text Reader

Abstract

This application relates to the field of power supply fault technology, and provides a method and system for diagnosing power supply faults in AI servers. The method includes: determining capacitor aging characteristic information of capacitors and connector aging characteristic information of connectors in the power supply based on multiple steady-state voltage values; determining the capacitor aging index based on the capacitor aging characteristic information; determining the connector aging index based on the connector aging characteristic information; and determining the fault diagnosis result of the AI ​​server's power supply based on the capacitor aging index and the connector aging index. This solution can improve the efficiency of AI server power supply fault diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of power supply failure technology, and in particular to a method and system for diagnosing power supply failures in AI servers. Background Technology

[0002] In routine data center operations and maintenance, diagnosing hidden faults in AI server power systems typically relies on periodic inspections and manual trend analysis. Maintenance personnel periodically export historical monitoring data of the server power supply, plot long-term voltage trend charts, and visually inspect for persistent abnormal drift by combining this data with a single threshold specified in the maintenance manual.

[0003] However, in scenarios where AI servers operate under long-term high loads, the aging of internal capacitors in the power supply leads to a decrease in output capacity, while the connectors experience increased contact resistance due to thermal expansion and contraction and mechanical vibration. These two components often degrade simultaneously at different rates. In the output voltage trend graph, the increased ripple and slight voltage drop caused by capacitor aging, along with the increased load-bearing voltage drop caused by connector degradation, collectively create a blurred, accelerating downward trend. Since maintenance personnel can only observe a single output voltage trend curve, when the degradation rates of these two components differ and both contribute to the long-term voltage trend in the same direction, the single trend curve cannot be broken down into the individual contributions of different degraded components. This makes it difficult for the single threshold rule based on total voltage offset in the maintenance manual to accurately distinguish the source of the fault. This compound degradation effect makes the attribution of latent faults ambiguous, often leading to delays in handling potential risks. Summary of the Invention

[0004] This application provides a method and system for diagnosing power supply faults in AI servers, aiming to solve the problem that the combined voltage drop trend caused by capacitor aging and connector degradation in AI server power supply systems is difficult to analyze, thus making it impossible to accurately diagnose hidden faults.

[0005] The technical solution of this application is as follows: In a first aspect, this application discloses a method for diagnosing power supply faults in an AI server, comprising: acquiring the steady-state voltage value of the power supply of the AI ​​server for each of multiple consecutive natural days; determining the capacitor aging characteristic information of the capacitors in the power supply and the connector aging characteristic information of the connectors in the power supply based on the multiple steady-state voltage values ​​of the power supply; determining the capacitor aging index of the capacitors based on the capacitor aging characteristic information of the capacitors; determining the connector aging index of the connectors based on the connector aging characteristic information of the connectors; and determining the fault diagnosis result of the power supply of the AI ​​server based on the capacitor aging index and the connector aging index of the connectors.

[0006] Optionally, obtaining the steady-state voltage value of the AI ​​server's power supply within each of multiple consecutive natural days includes: for each of the multiple natural days, obtaining the original voltage value sequence and the original current value sequence of the power supply within that natural day; determining whether a target time period exists within that natural day based on the original current value sequence; the target time period is a time period in which the ratio of the range of the original current values ​​to the average value of the original current values ​​within that natural day is less than a first preset ratio, and the duration is greater than or equal to a first preset duration; when a target time period exists within that natural day, using the average value of all original voltage values ​​corresponding to the target time period in the original voltage value sequence as the steady-state voltage value of the AI ​​server's power supply within that natural day; when no target time period exists within that natural day, obtaining the device importance index of the AI ​​server, and determining the steady-state voltage value of the AI ​​server's power supply within that natural day based on the device importance index, the original voltage value sequence of the power supply within that natural day, and the original current value sequence of the power supply.

[0007] Optionally, based on the device importance index, the original voltage value sequence of the power supply within a natural day, and the original current value sequence of the power supply, the steady-state voltage value of the AI ​​server's power supply within a natural day is determined, including: using the product of the device importance index and a preset current value coefficient as the initial current value step size; rounding up the initial current value step size to obtain the target current value step size; using the target current value step size as the step size, dividing the original current values ​​in the original current value sequence into multiple different current value intervals; counting the number of current values ​​in each current value interval; selecting the current value interval with the most current values ​​as the target current value interval; and using the median of the voltage values ​​corresponding to all current values ​​in the target current value interval as the steady-state voltage value of the AI ​​server's power supply within a natural day.

[0008] Optionally, the capacitor aging characteristic information of the capacitors in the power supply and the connector aging characteristic information of the connectors in the power supply are determined based on multiple steady-state voltage values ​​of the power supply, including: fitting an initial low-order polynomial curve of multiple steady-state voltage values ​​of the power supply based on the weighted least squares method; determining the residual between each steady-state voltage value of the power supply and the initial low-order polynomial curve; taking the steady-state voltage value with a residual greater than a preset residual threshold as the target steady-state voltage value; reducing the fitting weight of the target steady-state voltage value during the fitting of the initial low-order polynomial curve by a preset weight step size to obtain the updated fitting weight of the target steady-state voltage value; refitting the low-order polynomial curve of multiple steady-state voltage values ​​of the power supply based on the updated fitting weight of the target steady-state voltage value to obtain the target low-order polynomial curve of multiple steady-state voltage values ​​of the power supply; using the target low-order polynomial curve as the capacitor aging characteristic information of the capacitors; determining the residual sequence between the initial low-order polynomial curve and the target low-order polynomial curve, and using the residual sequence as the connector aging characteristic information of the connectors.

[0009] Optionally, the capacitor aging index is determined based on the capacitor's aging characteristic information, including: obtaining the reference health voltage of the AI ​​server's power supply; using the difference between the terminal value of the target low-order polynomial curve and the reference health voltage as the long-term cumulative offset; performing a first-order difference operation on the target low-order polynomial curve to obtain the average slope of the target low-order polynomial curve, and using the average slope as the long-term degradation rate; dividing the long-term cumulative offset by the capacitor's preset failure tolerance limit voltage difference to obtain the offset percentage; dividing the long-term degradation rate by the capacitor's preset rate reference constant to obtain the rate hazard percentage; and using the weighted sum of the offset percentage and the rate hazard percentage as the capacitor's aging index.

[0010] Optionally, determining the connector aging index based on the connector aging characteristic information includes: acquiring maintenance record information for each natural day among multiple natural days; the maintenance record information is used to indicate whether maintenance operations exist on a natural day; for each natural day, when the maintenance record information indicates that maintenance operations exist on that natural day, that natural day is taken as the target natural day; based on linear interpolation, determining the new residual corresponding to the target natural day in the residual sequence, replacing the residual corresponding to the target natural day in the residual sequence with the new residual of the target natural day to obtain the target residual sequence; and determining the connector aging index based on the target residual sequence.

[0011] Optionally, determining the connector aging index based on the target residual sequence includes: determining the standard deviation of the target residual sequence; using the ratio of the standard deviation of the target residual sequence to the preset normal residual standard deviation as the fluctuation amplification factor; performing quadratic polynomial fitting on the target residual sequence to obtain the fitting result, and extracting the quadratic term coefficients from the fitting result; using the product of the fluctuation amplification factor and the preset fluctuation amplification factor as the fluctuation index; using the product of the quadratic term coefficient and the preset thermal runaway factor as the thermal runaway index; and using the weighted sum of the fluctuation index and the thermal runaway index as the connector aging index of the connector.

[0012] Optionally, the fault diagnosis result of the AI ​​server's power supply is determined based on the capacitor aging index and the connector aging index, including: obtaining the ambient temperature sequence of the AI ​​server's power supply environment for each of multiple consecutive natural days; determining the average temperature rise gradient value of the ambient temperature sequence; determining whether the average temperature rise gradient value is less than a preset temperature rise gradient value; when the average temperature rise gradient value is less than the preset temperature rise gradient value, determining whether the capacitor aging index of the capacitor is greater than a preset capacitor aging index threshold; and determining whether the connector aging index of the connector is greater than a preset connector aging index threshold; when the capacitor aging index of the capacitor is greater than the preset capacitor aging index threshold, or the connector aging index of the connector is greater than the preset connector aging index threshold, the fault diagnosis result of the AI ​​server's power supply is determined to be an aging fault; otherwise, the fault diagnosis result of the AI ​​server's power supply is determined to be an aging fault; when the average temperature rise gradient value is greater than or equal to the preset temperature rise gradient value, the fault diagnosis result of the AI ​​server's power supply is determined based on the average temperature rise gradient value, the capacitor aging index, and the connector aging index.

[0013] Optionally, the fault diagnosis result of the AI ​​server's power supply is determined based on the average temperature rise gradient value, the capacitor's aging index, and the connector's aging index. This includes: using the product of the average temperature rise gradient value and a preset capacitor aging coefficient as the capacitor adjustment step size; using the product of the average temperature rise gradient value and a preset connector aging coefficient as the connector adjustment step size; using the difference between a preset capacitor aging index threshold and the capacitor adjustment step size as the target capacitor aging index threshold; using the difference between a preset connector aging index threshold and the connector adjustment step size as the target connector aging index threshold; determining whether the capacitor's aging index is greater than the target capacitor aging index threshold; determining whether the connector's aging index is greater than the target connector aging index threshold; if the capacitor's aging index is greater than the target capacitor aging index threshold, or the connector's aging index is greater than the target connector aging index threshold, the fault diagnosis result of the AI ​​server's power supply is determined to be an aging fault; otherwise, the fault diagnosis result of the AI ​​server's power supply is determined to be an aging fault.

[0014] Secondly, this application discloses an AI server power supply fault diagnosis system, comprising: an acquisition device and a processing device; the acquisition device is used to acquire the steady-state voltage value of the AI ​​server power supply in each of multiple consecutive natural days; the processing device is used to determine the capacitor aging characteristic information of the capacitors in the power supply and the connector aging characteristic information of the connectors in the power supply based on the multiple steady-state voltage values ​​of the power supply; the processing device is further used to determine the capacitor aging index of the capacitors based on the capacitor aging characteristic information of the capacitors; the processing device is further used to determine the connector aging index of the connectors based on the connector aging characteristic information of the connectors; the processing device is further used to determine the fault diagnosis result of the AI ​​server power supply based on the capacitor aging index and the connector aging index of the connectors. Beneficial effects

[0015] This application, through multi-dimensional analysis of the steady-state voltage values ​​of AI server power supplies, utilizes weighted least squares fitting and residual sequence analysis to successfully decouple capacitor aging characteristics from connector contact defects within a single voltage trend. This method solves the problem of ambiguity in fault diagnosis caused by composite degradation in existing technologies. By using quantified capacitor and connector aging indices combined with environmental temperature rise gradients for dynamic threshold determination, it achieves accurate identification and early warning of latent power supply faults, effectively improving the operational efficiency of data centers and the operational stability of server systems. Attached Figure Description

[0016] Figure 1 A flowchart illustrating an AI server power supply fault diagnosis method provided in this application; Figure 2 A flowchart illustrating an AI server power supply fault diagnosis method provided in this application; Figure 3 This application provides an architectural diagram of an AI server power supply fault diagnosis system. Detailed Implementation

[0017] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following drawings indicate similar items; therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0018] like Figure 1 As shown, the technical solution disclosed in this embodiment is as follows: S101. Obtain the steady-state voltage value of the AI ​​server's power supply for each of the multiple consecutive natural days.

[0019] S102. Determine the capacitor aging characteristics of the capacitors in the power supply and the connector aging characteristics of the connectors in the power supply based on multiple steady-state voltage values ​​of the power supply.

[0020] S103. Determine the capacitor aging index based on the capacitor aging characteristic information.

[0021] S104. Determine the connector aging index based on the connector aging characteristic information of the connector.

[0022] S105. Determine the fault diagnosis result of the AI ​​server's power supply based on the capacitor aging index and the connector aging index.

[0023] The following will combine Figure 2 The proposed solution will be explained in detail.

[0024] The overall operational logic of the above technical solution is as follows: First, using "natural days" as the time span unit, the steady-state voltage values ​​of the AI ​​server power supply are continuously collected to form a long-term voltage observation sequence. Then, by decoupling the features of this long-term voltage observation sequence, independent feature information representing capacitor degradation and connector degradation is extracted. Next, this feature information is quantified into an intuitive aging index. Finally, the aging indices of these two dimensions are combined to output the final fault diagnosis result. The reason for using "natural days" as the organizational granularity is that power supply aging is a slow-variable evolution process. The steady-state voltage values ​​summarized by natural days can both preserve the long-term degradation trend and avoid introducing a large amount of operational noise when excessively high-frequency sampling is directly involved in the diagnosis. The sequence formed by multiple consecutive natural days is essentially a continuous observation record of the health status of the same power supply on different dates. Subsequent feature decoupling, index mapping, and fault attribution judgment are all based on this record.

[0025] The core technical approach of "determining capacitor aging characteristics and connector aging characteristics in a power supply based on multiple steady-state voltage values" aims to separate the complex voltage change trends. In practice, frequency domain separation or trend / fluctuation separation logic can be employed. For example, a time series of steady-state voltage values ​​from multiple consecutive natural days is constructed. Data alignment is first performed according to date order, and missing dates are marked. For dates with missing values, interpolation with nearby dates, retaining the missing value and skipping it for local analysis, or directly excluding the data segment from the current diagnostic round if the missing percentage exceeds a preset range, can be used. After data processing, a low-pass filtering algorithm is used to extract the long-term, slowly declining baseline from the time series. This baseline data is used as capacitor aging characteristic information, as capacitor aging typically manifests as a long-term, irreversible, slow voltage decrease. The low-pass filtering is not limited to a specific algorithm; methods such as moving averages, exponential smoothing, and trend-fitting correction can be used, as long as the slow, continuous change components spanning multiple natural days are preserved. Meanwhile, high-frequency fluctuations or variance changes that deviate from the baseline in the time series are extracted and used as connector aging characteristic information, because poor connector contact often shows more obvious voltage fluctuation amplification under different loads.

[0026] After acquiring the aforementioned feature information, the system will convert the capacitor aging feature information and connector aging feature information into values ​​ranging from 0 to 100, respectively, using a preset mapping function or normalization algorithm. These values ​​are the capacitor aging index and the connector aging index. This "preset" is not arbitrarily given but can be established using historical samples. Specifically, the system can collect the steady-state voltage sequence of AI server power supplies from the period prior to the failure, based on samples from AI servers that have undergone disassembly or repair confirmation. This sequence is then categorized according to the actual repair results, such as "primarily capacitor aging," "primarily connector aging," "dual-component composite aging," and "no obvious aging." Subsequently, the system statistically analyzes the degree of long-term baseline decline, the degree of fluctuation enhancement, and the voltage drop dispersion within similar load ranges for different categories of samples, forming a mapping rule from feature quantities to index ranges. This mapping rule can be implemented using a segmented normalization table, a lookup table, or a hierarchical scoring method.

[0027] Through the aforementioned overall technical solution, this application breaks through the limitations of traditional operation and maintenance that relies solely on visual judgment using a single voltage trend curve. By decoupling the aging characteristics of capacitors and connectors from long-term steady-state voltage values ​​and quantifying them into independent aging indices, this solution can clearly reveal the degradation evolution process of different components within the power supply. This decoupling analysis mechanism effectively solves the problem of compound trend interference caused by the combined effect of slow capacitor degradation and accelerated connector contact resistance degradation, enabling the system to accurately determine the specific attribution of the output voltage decline trend, thereby achieving precise location and early warning of latent faults. Further analysis of its technical effects shows that if only the total voltage drop is observed, the long-term slow decline and short-term fluctuation amplification are superimposed on the same curve, making fault attribution easy to confuse; however, in this solution, the long-term baseline is extracted first, and then the discrete components deviating from the baseline are analyzed, which is equivalent to observing "slow changes" and "fast changes" in layers. Capacitor aging mainly affects long-term support capability, and therefore is more concentrated in the baseline layer; connector aging is more likely to amplify voltage drop differences when the load changes, and therefore is more concentrated in the fluctuation layer. It is precisely because the two types of degradation mechanisms differ in their temporal manifestations that decoupling based on single voltage data becomes possible, ultimately leading to clearer fault attribution results.

[0028] In some preferred embodiments, obtaining the steady-state voltage value of the AI ​​server's power supply within each of multiple consecutive natural days includes: for each of the multiple natural days, obtaining the original voltage value sequence and the original current value sequence of the power supply within that natural day; determining whether a target time period exists within that natural day based on the original current value sequence; the target time period is a time period in which the ratio of the range of the original current values ​​to the average value of the original current values ​​within that natural day is less than a first preset ratio, and the duration is greater than or equal to a first preset duration; when a target time period exists within that natural day, taking the average of all original voltage values ​​corresponding to the target time period in the original voltage value sequence as the steady-state voltage value of the AI ​​server's power supply within that natural day; when no target time period exists within that natural day, obtaining the device importance index of the AI ​​server, and determining the steady-state voltage value of the AI ​​server's power supply within that natural day based on the device importance index, the original voltage value sequence of the power supply within that natural day, and the original current value sequence of the power supply.

[0029] Specifically, the system first synchronously acquires high-frequency raw voltage and current value sequences through underlying sensors. "Synchronous acquisition" here means that voltage and current values ​​are recorded at the same sampling time or within the same extremely short sampling period to ensure accurate establishment of the "voltage performance corresponding to a certain current state" during subsequent analysis. The raw voltage and current value sequences can be obtained in the following ways: voltage and current detection channels are set up on the AI ​​server's power output side; the detection results are sent to the BMC after analog-to-digital conversion; the BMC adds a unified timestamp to each set of sampled data and writes it to a cache file, time-series database, or monitoring platform according to natural days. If the server uses rack-level centralized power supply, the rack power management unit can also collect the data and distribute it to the corresponding server identifier. To ensure data availability, the system can also perform a basic cleaning before writing, including removing abnormal values ​​that clearly exceed the power supply's rated output range, removing duplicate timestamp records, and filling in gaps caused by short-term communication interruptions.

[0030] Subsequently, the system slides a window on the timeline to perform a stability scan on the original current value sequence, searching for a "target time period." The criteria for determining this target time period are extremely stringent: first, the difference between the maximum and minimum current values ​​within this time period, divided by the average current value for that period, must be less than a first preset ratio; second, this stable state must last for more than a first preset duration. The first preset ratio and first preset duration can be obtained statistically from historical data based on the data center's business fluctuation characteristics, server type, and sampling period. A feasible approach is to first select a batch of known stable date samples, statistically analyze the dispersion of time periods with relatively small current fluctuations within these samples, and then select a threshold that can effectively distinguish between "stable load" and "frequent load switching" as the first preset ratio; simultaneously, combining the time required for power supply thermal state establishment and the distribution of business task durations, select a duration that covers stable operating segments as the first preset duration. This approach yields thresholds that better reflect the actual deployment environment, rather than being a fixed setting detached from the scenario.

[0031] When the system successfully identifies a target time period with extremely stable current within a given calendar day, it indicates that the power supply's output voltage is unaffected by drastic load changes. The system then averages all the raw voltage values ​​within that target time period to determine the steady-state voltage for that calendar day. If multiple target time periods meet the criteria within the same calendar day, the longest-lasting target time period can be selected, or the average voltage of each target time period can be calculated separately and then aggregated. Prioritizing the longer duration and smaller current fluctuations typically yields a more stable steady-state voltage representation. However, for some core AI servers operating at full capacity throughout the day with frequent task switching, a stable time period meeting the above criteria may not be found. In this case, the system triggers a backup mechanism, retrieving the "equipment importance index" pre-configured in the AI ​​server's asset management database. This index, combined with the entire day's raw voltage and current value sequences, is used to calculate the steady-state voltage value through a more sophisticated statistical algorithm.

[0032] Through this specific implementation plan, the system establishes an adaptive steady-state voltage extraction logic. When the load has a stable period, a clean observation window is locked by strictly defining the ratio of current range to mean, eliminating the main interference of dynamic load on voltage sampling. When the load fluctuates drastically throughout the day, an important equipment index is introduced as an adjustment factor for degradation processing. This mechanism is effective because the core objective of steady-state voltage is not simply to obtain the average voltage of a certain day, but to find the voltage benchmark that best represents the state of the power supply itself. Current stability scanning first filters out periods with large load disturbances at the source, making the remaining voltage closer to the healthy state of the power supply itself. When such an ideal period does not exist, statistical methods are used to find the most representative load range of the day, thereby avoiding the results being dominated by extreme peaks or valleys. As a result, regardless of the complexity of the AI ​​server's business conditions, high-quality and high-confidence steady-state voltage values ​​can be extracted, providing a data foundation for subsequent aging feature decoupling.

[0033] Furthermore, for the situation where there is no target time period within a natural day, in practical applications, the steady-state voltage value of the AI ​​server's power supply within a natural day is determined based on the equipment importance index, the original voltage value sequence of the power supply within a natural day, and the original current value sequence of the power supply. This includes: using the product of the equipment importance index and a preset current value coefficient as the initial current value step size; rounding up the initial current value step size to obtain the target current value step size; using the target current value step size as the step size, dividing the original current values ​​in the original current value sequence into multiple different current value intervals; counting the number of current values ​​in each current value interval; selecting the current value interval with the most current values ​​as the target current value interval; and using the median of the voltage values ​​corresponding to all current values ​​in the target current value interval as the steady-state voltage value of the AI ​​server's power supply within a natural day.

[0034] In practical implementation, the system first uses equipment importance indices to determine the granularity of statistical analysis. Equipment importance indices can be pre-stored in an asset management database, configuration management database, or operations and maintenance orchestration platform. They can be manually configured or written based on hierarchical classifications according to business level, service targets, redundancy capabilities, and the scope of downtime impact. For example, AI servers undertaking core training tasks but lacking immediate redundancy switching capabilities can be assigned a higher equipment importance index; AI servers undertaking testing tasks or possessing sufficient redundancy replacement capabilities can be assigned a lower equipment importance index.

[0035] Subsequently, the system multiplies the device importance index by a preset current value coefficient to obtain the initial current value step size; then, it rounds up the initial current value step size to obtain the target current value step size. Rounding up facilitates the division of continuous current values ​​into clearly defined and easily implemented discrete intervals, reducing the complexity of interval boundaries caused by decimals in the step size. Next, using the target current value step size as the step size, the system divides the raw current values ​​within the range of low to high loads within that natural day into multiple continuous current value intervals. The division can start from the minimum current value observed within that natural day or from the common low load boundary of the device, as long as all raw current values ​​fall within a certain interval. Then, the system traverses the entire day's raw current value sequence, assigning each current sampling point to its corresponding current value interval and counting the total number of current data points clustered within each interval. The interval containing the most data points is the target current value interval, representing the most frequent and typical load condition for the AI ​​server during that day. Here, "most numerous" reflects the dominant operating state in terms of time proportion, rather than the instantaneous peak state, and is therefore more suitable as the basis for steady-state voltage extraction.

[0036] Finally, the system extracts the original voltage values ​​corresponding to all current points falling within the target current value range at the same time, sorts these voltage values ​​by size, and takes the median as the steady-state voltage value for that day. Using the median instead of the average value has a clear technical consideration: in high-frequency fluctuation scenarios, even when limited to the target current value range, there may still be a small number of abnormal voltage points caused by transient switching, sampling jitter, local contact disturbances, or monitoring noise; if the average value is directly taken, these points with large deviations will still affect the result, while the median is not sensitive to a small number of abnormal points and better reflects the typical voltage level within the range.

[0037] Through this specific implementation scheme, the system cleverly transforms chaotic dynamic load data into a structured probability distribution model. By dynamically generating target current value step sizes using device importance indices, it allows devices of varying importance to obtain statistical granularity that matches their business characteristics. Finding the target current value range with the most data points essentially locks into the server's most typical operating state throughout the day; and using the median voltage within this range, rather than the average, further eliminates potential transient voltage spikes. This combined mechanism of dynamic step-size interval statistics and median extraction can accurately anchor the most representative voltage benchmark even under extreme operating conditions with no stable load periods, greatly improving the robustness of power supply fault diagnosis in complex business scenarios.

[0038] In some preferred embodiments, determining the capacitor aging characteristics of capacitors and the connector aging characteristics of connectors in the power supply based on multiple steady-state voltage values ​​includes: fitting initial low-order polynomial curves of multiple steady-state voltage values ​​of the power supply using a weighted least squares method; determining the residuals between each steady-state voltage value of the power supply and the initial low-order polynomial curves; using steady-state voltage values ​​with residuals greater than a preset residual threshold as target steady-state voltage values; reducing the fitting weights of the target steady-state voltage values ​​during the fitting of the initial low-order polynomial curves by a preset weight step size to obtain updated fitting weights for the target steady-state voltage values; refitting the low-order polynomial curves of multiple steady-state voltage values ​​of the power supply based on the updated fitting weights for the target steady-state voltage values ​​to obtain target low-order polynomial curves for the multiple steady-state voltage values ​​of the power supply; using the target low-order polynomial curves as capacitor aging characteristics; determining the residual sequence between the initial low-order polynomial curves and the target low-order polynomial curves, and using the residual sequence as connector aging characteristics.

[0039] The core of this specific implementation plan lies in gradually reducing the impact of abnormal fluctuations on baseline fitting through iterative weight reduction. First, the system uses weighted least squares to initially fit the steady-state voltage values ​​over multiple consecutive natural days. In engineering implementation, a one-to-one daily index can be established for multiple steady-state voltage values ​​according to time sequence, such as day 1, day 2, up to day N. The daily index is used as the horizontal axis for fitting, and the steady-state voltage value is used as the vertical axis. In the initial state, all steady-state voltage data points can be assigned the same initial weight (e.g., all weights are 1), fitting an initial low-order polynomial curve. The low-order polynomial here is usually chosen as first-order (linear) or second-order (parabolic) because the physical aging process of capacitors exhibits a smooth and slow monotonous decrease over a timescale of months or even years. High-order polynomials are prone to overfitting, incorrectly absorbing components belonging to connector fluctuations into the trend line. To more clearly select the fitting order, first-order curves can be prioritized for scenarios where voltage drops are mostly uniform, while second-order curves are used for scenarios where there is slight accelerated degradation in the later stages of aging but the overall trend remains smooth. If unreasonable reverse bending is observed in local areas after using a second-order curve, it can revert to a first-order curve to ensure that the target low-order polynomial curve is consistent with the physical law of slow capacitor degradation. Subsequently, the system calculates the vertical difference between each steady-state voltage value and the corresponding point of the initial low-order polynomial curve, i.e., the residual. The residual reflects the degree to which the voltage deviates from the long-term trend on that natural day. The residual can retain directional information to distinguish deviations above or below the trend line, and can also record its absolute deviation magnitude for subsequent determination of whether it exceeds a preset residual threshold.

[0040] Next, the system compares these residuals with a preset residual threshold. The preset residual threshold can be determined by statistically analyzing the voltage residual distribution of historical healthy power supplies under normal load fluctuations and taking the upper limit of its 95% confidence interval. Further, the preset residual threshold can be obtained as follows: First, select power supply samples of the same model, power rating, and connection structure, confirmed to have no fault records, and collect their steady-state voltage values ​​under standard computer room temperature and humidity conditions, typical load ranges, and over multiple consecutive natural days; then, use the same initial low-order polynomial fitting method as online to obtain the residual distribution of healthy samples, and set the upper limit that the vast majority of healthy residuals will not exceed as the preset residual threshold. If different power supply models have different voltage ratings, it is preferable to establish their own threshold tables separately to avoid incorrectly applying the normal fluctuation range of one model to another. When the residual of a certain steady-state voltage value is greater than the preset residual threshold, it indicates that this point is very likely an abnormal fluctuation point caused by poor connector contact, and the system marks it as the target steady-state voltage value. To prevent these target steady-state voltage values ​​from skewing the true capacitor aging baseline, the system reduces their weights during the fitting process by a preset weight step size (e.g., by 0.5 or by half) to obtain updated fitting weights; while for steady-state voltage values ​​with residuals within the normal range, their weights remain unchanged.

[0041] In terms of overall working principle, this specific implementation scheme achieves robust separation of trends and fluctuations through a closed-loop mechanism of "initial fitting - residual evaluation - weight reduction - refitting". The initial low-order polynomial curve provides a rough benchmark, enabling the system to identify which points are abnormal fluctuation points deviating from the benchmark. By reducing the weights of these abnormal fluctuation points and refitting, the system forcibly weakens the "gravitational pull" of the connector's violent fluctuations on the long-term trend of the capacitance, so that the final target low-order polynomial curve can closely fit those stable data points that truly reflect the state of the capacitance. This mechanism effectively solves the problem of feature distortion caused by extreme fluctuation points pulling the fitted curve as mentioned above. Its unique technical effect is that, in complex degradation scenarios, even if the voltage fluctuations caused by connector aging are extremely violent and frequent, this solution can still act like a "filter" to accurately intercept these high-frequency violent fluctuations outside the trend fitting process, ensuring that the extracted capacitor aging baseline is highly pure and undistorted; at the same time, the intercepted residual sequence also retains the most complete fluctuation energy, making the subsequent quantitative assessment of connector aging more sensitive and accurate, and greatly improving the robustness and accuracy of dual component feature decoupling.

[0042] Optionally, the capacitor aging index is determined based on the capacitor's aging characteristic information, including: obtaining the reference health voltage of the AI ​​server's power supply; using the difference between the terminal value of the target low-order polynomial curve and the reference health voltage as the long-term cumulative offset; performing a first-order difference operation on the target low-order polynomial curve to obtain the average slope of the target low-order polynomial curve, and using the average slope as the long-term degradation rate; dividing the long-term cumulative offset by the capacitor's preset failure tolerance limit voltage difference to obtain the offset percentage; dividing the long-term degradation rate by the capacitor's preset rate reference constant to obtain the rate hazard percentage; and using the weighted sum of the offset percentage and the rate hazard percentage as the capacitor's aging index.

[0043] This specific implementation plan aims to quantify the aging degree of capacitors from two dimensions: "depth" and "speed." First, the system acquires the reference health voltage of the AI ​​server power supply. This voltage is typically the steady-state voltage reference value measured under standard load when the power supply is in good condition during its initial factory testing or first online operation. It represents the ideal output level when the capacitors have not undergone any aging. There are several feasible paths to obtain this reference health voltage: one approach is during the power supply's factory testing phase, where a testing platform collects several sets of steady-state voltages under specified input conditions, rated output conditions, and a stable temperature environment, and then writes these values ​​into the device file. Another approach is during the break-in period after the server's initial online operation, selecting several days with no alarms, no abnormal maintenance, and stable load, extracting the steady-state voltage values ​​of these days, and forming an initial health reference. For devices that have been running for a period of time but lack factory reference records, the reference health voltage of healthy devices from the same batch and model can be used as a substitute reference, and its source can be marked in the file. Next, the system extracts the last data point (i.e., the terminal value) of the target low-order polynomial curve on the time axis. This terminal value represents the true voltage baseline level of the power supply at the end of the current observation period after long-term smoothing. Subtracting this terminal value from the reference healthy voltage, the difference is the long-term cumulative offset, which intuitively reflects the total voltage drop caused by the capacitance decay since the capacitor was put into operation.

[0044] Here, a unified directional definition of "reference healthy voltage minus the final value" is preferred to obtain a positive characterization of aging depth under voltage drop scenarios, avoiding confusion in subsequent comparisons caused by different difference directions used by different devices. Subsequently, the system performs a first-order difference operation on the target low-order polynomial curve. Since the curve is a low-order polynomial, its first-order difference (or derivative) reflects the slope of the curve. The system extracts its average slope as the long-term degradation rate, which characterizes the speed of the capacitor's current performance degradation. To make the long-term degradation rate more stable, the average change level over the entire observation period or the average change level within a recent window can be used; the former is suitable for reflecting the overall degradation rhythm, while the latter is more suitable for emphasizing whether the recent stage of accelerated degradation has been entered. In actual deployment, the system configuration can be set to use either full-cycle mode or sliding window mode. To unify these two indicators with different physical dimensions into the same evaluation system, the system introduces a preset failure tolerance limit voltage difference and a preset rate reference constant. The preset failure tolerance limit voltage difference refers to the maximum voltage drop allowed in the power supply design specifications. If this drop is exceeded, the capacitor is considered to have completely failed. The preset rate reference constant is the critical rate of decline that indicates the capacitor is about to enter the accelerated decay period, which is statistically derived from historical failure samples.

[0045] Furthermore, these two preset values ​​can be established as follows: First, collect life test data, repair fault data, and long-term online monitoring data of the same model of power supply. Then, based on the voltage drop depth and rate distribution before the fault occurs, select reference values ​​that can better distinguish between the three states of "normal degradation," "needs attention," and "high-risk failure eve," and solidify them into the parameter table for the corresponding model. If necessary, the parameter files can be further subdivided according to ambient temperature range and common load rate range. The system divides the long-term cumulative offset by the preset failure tolerance limit voltage difference to obtain the offset percentage, which is used to measure the progress of the capacitor on the "life milestone"; it divides the long-term degradation rate by the preset rate reference constant to obtain the rate danger percentage, which is used to measure the current "acceleration towards failure" of the capacitor. Finally, the system assigns corresponding weights to these two percentages (e.g., offset accounts for 60%, rate danger accounts for 40%), calculates their weighted sum, and finally outputs a comprehensive value between 0 and 100, which is the capacitor aging index. The weights can be preset according to the operation and maintenance strategy: if more attention is paid to the risk of rapid deterioration, the weight of the rate of danger percentage can be appropriately increased; if more attention is paid to the degree of lifespan depletion, the weight of the deviation percentage can be appropriately increased. To ensure the traceability of configuration sources, the weights are preferably recorded in the equipment health assessment strategy table and associated with the equipment model.

[0046] In terms of working principle, this specific implementation scheme captures the static cumulative results and dynamic evolution trend of capacitor aging by extracting the terminal values ​​and average slope of the target low-order polynomial curve. By dividing by their respective physical limits or empirical benchmark constants, the absolute voltage and rate values ​​are transformed into dimensionless relative danger percentages, thereby eliminating the evaluation bias caused by the initial voltage differences of different power supply models. Finally, through weighted summation, the two dimensions of risk are merged into a single intuitive index. This mechanism effectively solves the problem of missed or false judgments that are easily caused by single-dimensional evaluation, as mentioned above. Its unique technical effect is that in application scenarios where capacitors degrade slowly, this scheme can not only accurately assess the "severity" of the capacitor (cumulative offset) but also keenly capture the "rapid deterioration" (degradation rate). This two-dimensional fusion quantification method enables the system to identify high-risk capacitors showing accelerated deterioration trends before the capacitor voltage falls below the hard alarm threshold, thus significantly advancing the fault warning time window and giving data center maintenance personnel more time for spare parts procurement and planned downtime replacement. Further analysis reveals that the key to its technical effectiveness lies in the fact that the cumulative offset reflects the lifespan already consumed, while the degradation rate reflects the intensity of ongoing changes. These correspond to risk information at two different levels: "outcome" and "trend." Focusing solely on the outcome can easily overlook situations where the recent rate of deterioration has sharply increased; focusing only on the trend can easily overlook situations where the system is nearing failure but short-term changes are slowing down. By combining these two aspects, the system can identify both "chronic high risk" and "rapid deterioration risk," thus better meeting the actual needs of preventative maintenance in data centers.

[0047] In practical applications, the connector aging index is determined based on the connector aging characteristic information, including: obtaining maintenance record information for each natural day across multiple natural days; the maintenance record information is used to indicate whether maintenance operations exist on a natural day; for each natural day, if the maintenance record information indicates that maintenance operations exist on that natural day, that natural day is taken as the target natural day; based on linear interpolation, a new residual corresponding to the target natural day is determined in the residual sequence, and the residual corresponding to the target natural day in the residual sequence is replaced with the new residual of the target natural day to obtain the target residual sequence; the connector aging index is determined based on the target residual sequence.

[0048] The core of this specific implementation plan lies in utilizing external operational context information to purify the underlying sensor data. First, the system interfaces with the data center's IT Service Management System (ITSM) or electronic operational log database to obtain maintenance records for multiple calendar days corresponding to the AI ​​server. These records contain explicit instructions on whether engineers performed physical operations such as unpacking, cable plugging / unplugging, or hardware upgrades on the server each day. Specifically, this can be achieved by: reading the device number, maintenance time, and maintenance type fields from the work order system via an interface call; or by having operational personnel manually enter the information into the electronic logs, which are then periodically captured and converted into structured records by the monitoring platform. To ensure a one-to-one correspondence between maintenance records and residual sequences, the system prioritizes standardizing time zones and time formats before merging by calendar day; for maintenance work orders spanning midnight, multiple days can be marked based on the calendar days covered by the actual impact time. Next, the system checks these records daily. Once it finds that a particular calendar day's maintenance records show that the aforementioned maintenance operations occurred on that day, the system marks that calendar day as the target calendar day.

[0049] Furthermore, if the maintenance log information contains more detailed operation types, the system can prioritize marking high-disruption operations such as power harness plugging / unplugging, connector re-plugging, module disassembly / assembly, and board addition as the target date. Operations that do not involve physical contact, such as software upgrades and remote reconfiguration, can be left unmarked or used only as a low-priority reference. Since the voltage residual on the target date is highly likely to be severely affected by human physical contact and cannot accurately reflect the natural contact state of the connector, the system decides to remove it and fill it appropriately. Specifically, the system uses linear interpolation to calculate a smooth transition estimate, i.e., the new residual, using the residual values ​​of adjacent normal dates before and after the target date that did not involve maintenance operations. The basic assumption of linear interpolation is that, without external human interference, the contact state of the connector evolves relatively continuously and smoothly over a short period. To ensure the reliability of the filled values, linear interpolation is preferred only when there are valid normal days before and after the target date. If the target date is at the beginning or end of the sequence, or if maintenance operations occur on multiple consecutive days, the nearest normal date can be searched forward or backward first. If necessary, the trend value of a longer window can be used for filling, and an "interpolation correction" label should be retained in the data tags for subsequent analysis modules to identify. The system then uses this calculated new residual to replace the abnormal residuals originally recorded for the target date in the residual sequence, thus obtaining a target residual sequence free from human interference. Finally, based on this purified target residual sequence, the system continues to perform subsequent connector aging index calculations. By retaining both versions of the residual sequence before and after correction, post-event auditing and effect retrospection are also supported, facilitating the analysis of the actual impact range of a maintenance operation on voltage fluctuations.

[0050] In terms of working principle, this specific implementation scheme establishes a context-aware anomaly removal mechanism by introducing cross-system data fusion (i.e., merging maintenance records with underlying voltage residuals). Maintenance record information acts as the trigger condition for the "filter," accurately locating data points that may be contaminated by non-aging factors; while linear interpolation acts as a "repair agent," removing contaminated data while maintaining the mathematical continuity and integrity of the time series, avoiding time axis misalignment in subsequent time series analysis algorithms (such as frequency analysis or polynomial fitting) due to direct deletion of data points. This mechanism effectively solves the problem of residual mutations caused by human maintenance operations, leading to false alarms. Its unique technical effect is that, in complex data center operation and maintenance environments, this solution gives the diagnostic system the ability to distinguish between "natural aging fluctuations" and "human physical interference." By accurately smoothing out false residual peaks caused by maintenance operations, the system ensures that the target residual sequence participating in the final evaluation purely reflects the physical degradation trajectory of the connector itself, thereby greatly reducing the false alarm rate of connector aging diagnosis and improving the reliability and practicality of diagnostic results in actual complex operation and maintenance scenarios. From a logical perspective, connector natural aging is typically continuous, gradual, and can persist across adjacent time periods, while disturbances caused by maintenance operations are usually sudden, localized, and verifiable by external work orders. This solution leverages this difference, using the presence or absence of maintenance operations as a crucial contextual basis for distinguishing between the two types of anomalies. Therefore, it can achieve higher-reliability data cleansing without disrupting the original sequence structure.

[0051] In a preferred embodiment of this application, determining the connector aging index of the connector based on the target residual sequence includes: determining the standard deviation of the target residual sequence based on the target residual sequence; using the ratio of the standard deviation of the target residual sequence to the preset normal residual standard deviation as the fluctuation amplification factor; performing quadratic polynomial fitting on the target residual sequence to obtain the fitting result, and extracting the quadratic term coefficients from the fitting result; using the product of the fluctuation amplification factor and the preset fluctuation amplification factor as the fluctuation index; using the product of the quadratic term coefficient and the preset thermal runaway factor as the thermal runaway index; and using the weighted sum of the fluctuation index and the thermal runaway index as the connector aging index of the connector.

[0052] This specific implementation scheme uses a combination of statistics and multinomial fitting to extract two key aging dimensions from the target residual sequence. First, the system calculates the standard deviation of the entire target residual sequence. Standard deviation is a classic statistic for measuring data dispersion; it objectively reflects the degree to which the voltage residual fluctuates around the baseline during this period. In practice, the target residual sequence can be read sequentially by natural day. First, the central level of the residuals within that time period is obtained, then the deviation of the residuals on each natural day from this central level is measured, ultimately forming the standard deviation. If the sequence length is too short, the standard deviation is easily affected by random factors. Therefore, it is preferable to output the connector aging index only when the minimum number of observation days is met. The minimum number of observation days can be set to a consecutive number of natural days based on operational practices. Next, the system divides the calculated standard deviation by the preset normal residual standard deviation. The preset normal residual standard deviation is based on the baseline value of the residual standard deviation measured under normal operating conditions from a large number of brand-new, healthy connectors of the same model.

[0053] The preset normal residual standard deviation can be obtained by establishing a healthy sample library. Specifically, this can be done by selecting devices that have not undergone maintenance, have not issued alarms, and have been operating stably for a long time from servers, power supplies, and connectors of the same model. Residual sequences for multiple natural days are collected under standard environments and typical business loads, their standard deviation distribution is calculated, and representative benchmark values ​​are selected and written into the parameter library. If ambient temperature or load level has a significant impact on residual fluctuations, different preset normal residual standard deviations can be maintained for different scenarios. The ratio of the two is the fluctuation amplification factor, which intuitively quantifies how much the current connector contact instability has deteriorated relative to the healthy state. Simultaneously, to capture the nonlinear divergence trend of the residual sequence on the time axis, the system performs quadratic polynomial fitting on the target residual sequence (i.e., fitting a curve of the form y=ax²+bx+c). In engineering implementation, the natural day number can be used as the time axis, and the target residual sequence as the fitting object to obtain an overall trend curve. In the fitting result, the coefficient of the quadratic term (i.e., parameter a) determines the opening direction and curvature of the curve. If the quadratic term coefficient is significantly positive, it indicates that the fluctuation amplitude of the residual exhibits an accelerating parabolic trend in the later stages of the observation period. This is a typical physical manifestation of "thermal runaway," where increased connector contact resistance leads to localized abnormal heating, which in turn causes accelerated deterioration of the contact condition. If the quadratic term coefficient is close to zero, it indicates that the residual change is closer to linear or stable fluctuation. If the quadratic term coefficient is negative, it indicates that the divergence trend is not obvious in the later stages, and the risk of thermal runaway is relatively low. To improve the interpretability of this indicator, the system can internally record a comparison chart of the quadratic polynomial fitting curve and the original target residual sequence for operation and maintenance verification.

[0054] Subsequently, the system multiplies the fluctuation amplification factor by a preset fluctuation amplification coefficient, mapping it to a standardized fluctuation index; it also multiplies the extracted quadratic term coefficient by a preset thermal runaway coefficient, mapping it to a standardized thermal runaway index. These two preset coefficients transform physical characteristics of different dimensions into a unified scoring scale. The method for obtaining these two preset coefficients can be based on historical sample calibration: a batch of connector samples confirmed to be in healthy, mildly aged, moderately aged, and severely aged states are collected, and their fluctuation amplification factor and quadratic term coefficient are calculated respectively. Then, combined with manual inspection results, temperature rise records, repair conclusions, or failure events, a more suitable preset fluctuation amplification coefficient and preset thermal runaway coefficient are calibrated to ensure that the output fluctuation index and thermal runaway index have a good correspondence with the actual risk level. Finally, the system performs a weighted sum of the fluctuation index and the thermal runaway index to obtain the final connector aging index. The weights can also be configured according to the focus of operation and maintenance: if you want to detect early signs of thermal runaway as early as possible, you can appropriately increase the weight of the thermal runaway index; if you are more concerned about the common contact instability problem, you can appropriately increase the weight of the volatility index.

[0055] In terms of working principle, this specific implementation scheme cleverly maps the physical mechanism of connector aging into mathematical characteristics. The calculation of standard deviation captures the global dispersion caused by unstable contact resistance, while the quadratic term coefficient in the quadratic polynomial fitting precisely locks the local accelerated divergence trend caused by heat accumulation. By transforming these two orthogonal mathematical characteristics into fluctuation exponents and thermal runaway exponents respectively and then weighting and fusing them, the system constructs an evaluation model that comprehensively covers the entire life cycle of connector degradation. This mechanism effectively solves the problem that the simple statistical methods mentioned above cannot capture complex degradation characteristics. Its unique technical effect is that in the diagnosis of connector aging, this scheme can not only identify "poor contact" (manifested as amplified fluctuation) in the slow degradation stage, but more importantly, it can keenly detect in advance that the connector is about to enter the dangerous "thermal runaway" stage (manifested as a significant increase in the quadratic term coefficient). This ability to capture nonlinear acceleration trends allows the system to issue highly forward-looking serious fault warnings before the connector burns out due to severe overheating or causes server downtime, thereby greatly improving the safety baseline of data center hardware operation. Further analysis of its mechanism reveals that simple fluctuation amplification often only indicates that the contact state is already unstable, but does not necessarily mean that it is about to fail rapidly. However, when the residual sequence simultaneously exhibits a significant trend of accelerated divergence in the later stages, it usually means that a vicious cycle of mutual reinforcement has formed between temperature rise, contact deterioration, and greater fluctuations. This scheme, by extracting these two features in parallel and scoring them uniformly, can clearly distinguish between "already unstable" and "accelerating deterioration" and incorporate them into the decision-making basis. Therefore, its early warning results are more suitable as a trigger condition for proactive maintenance of high-value AI servers in data centers.

[0056] After successfully quantifying the capacitor aging index and connector aging index through the aforementioned embodiments, the system needs to output the final fault diagnosis result based on these two indices. However, in actual data center environments, the aging performance of AI server power supplies is not entirely independent of the external environment. A continuous increase in ambient temperature accelerates the physicochemical degradation process of internal power supply components, making components already in a critical state more prone to substantial failure. If a fixed aging index threshold is used for diagnosis, in scenarios where ambient temperature continues to deteriorate, fault warnings may be delayed. Therefore, in a preferred embodiment of this application, determining the fault diagnosis result of the AI ​​server's power supply based on the capacitor aging index and the connector aging index includes: acquiring the ambient temperature sequence of the environment in which the AI ​​server's power supply is located for each of multiple consecutive natural days; determining the average temperature rise gradient value of the ambient temperature sequence; determining whether the average temperature rise gradient value is less than a preset temperature rise gradient value; when the average temperature rise gradient value is less than the preset temperature rise gradient value, determining whether the capacitor aging index is greater than a preset capacitor aging index threshold; and determining whether the connector aging index is greater than a preset connector aging index threshold; when the capacitor aging index is greater than the preset capacitor aging index threshold, or the connector aging index is greater than the preset connector aging index threshold, determining that the fault diagnosis result of the AI ​​server's power supply is that the power supply has an aging fault; otherwise, determining that the fault diagnosis result of the AI ​​server's power supply is that the power supply does not have an aging fault; when the average temperature rise gradient value is greater than or equal to the preset temperature rise gradient value, determining the fault diagnosis result of the AI ​​server's power supply based on the average temperature rise gradient value, the capacitor aging index, and the connector aging index.

[0057] This specific implementation plan introduces ambient temperature as a context variable to construct a dynamic diagnostic mechanism. Specifically, the system first acquires the ambient temperature sequence of the AI ​​server's power supply environment for each of several consecutive natural days. This ambient temperature sequence can be obtained by using temperature sensors at the AI ​​server's front panel air inlet, thermistors inside the power module, or temperature and humidity sensors in the rack's cold aisle, collecting temperature data at a fixed frequency, and calculating the daily average temperature or daily maximum temperature for each natural day, thus forming a temperature observation sequence aligned with the steady-state voltage value on the time axis. Next, the system determines the average temperature rise gradient value of this ambient temperature sequence. The average temperature rise gradient value reflects the overall upward trend and deterioration rate of the ambient temperature throughout the observation period. In engineering implementation, the least squares method can be used to linearly fit the ambient temperature sequence, and the slope of the fitted line can be extracted as the average temperature rise gradient value; or the difference between the first and last temperatures of the sequence can be calculated and divided by the number of observation days. Subsequently, the system compares the calculated average temperature rise gradient value with a preset temperature rise gradient value. The preset temperature rise gradient value is the critical threshold that distinguishes between normal temperature fluctuations and abnormal continuous temperature rises in the computer room. It can be set by statistically analyzing the historical extreme values ​​of temperature fluctuations during normal operation of the computer room air conditioning system.

[0058] When the average temperature rise gradient value is less than the preset temperature rise gradient value, it indicates that the ambient temperature of the current AI server is relatively stable and there is no obvious trend of continuous deterioration. At this time, the system adopts the conventional static threshold judgment logic, that is, it judges whether the capacitor aging index is greater than the preset capacitor aging index threshold and whether the connector aging index is greater than the preset connector aging index threshold. These two preset thresholds are baseline safety red lines calibrated based on historical fault data of the same type of power supply under standard ambient temperature. As long as any one of the aging indices exceeds the corresponding preset threshold, the system determines that the fault diagnosis result of the AI ​​server's power supply is that the power supply has an aging fault, and can further output a detailed conclusion of "capacitor aging fault" or "connector aging fault" according to the specific index type that exceeds the limit; if neither exceeds the limit, it is determined that the power supply does not have an aging fault. Conversely, when the average temperature rise gradient value is greater than or equal to the preset temperature rise gradient value, it indicates that there may be hidden dangers in the data center heat dissipation system, or that the server has been in a high heat load state for a long time, at which point the tolerance of the components has decreased. The system will trigger a dynamic diagnosis branch, and comprehensively determine the fault diagnosis result based on the average temperature rise gradient value, the capacitor aging index, and the connector aging index.

[0059] This specific implementation plan overcomes the limitations of traditional diagnostic models that are "environment-independent." Its unique technical advantage lies in introducing the average temperature rise gradient as an indicator of environmental deterioration, enabling the system to adaptively perceive deteriorating external operating conditions. When the ambient temperature is stable, a standard threshold is used to ensure diagnostic stability and control the false alarm rate; while when the ambient temperature continues to rise, it promptly switches to dynamic diagnostic logic. This adaptive switching mechanism based on environmental context makes the fault diagnosis results more closely reflect the actual withstand capabilities of power components in real physical environments, effectively avoiding missed alarms or delayed warnings caused by rigid evaluation standards under harsh operating conditions.

[0060] To address the situation where the average temperature rise gradient value is greater than or equal to a preset temperature rise gradient value, and in order to achieve reasonable dynamic adjustment of the aging index threshold, in a preferred embodiment of this application, the fault diagnosis result of the AI ​​server's power supply is determined based on the average temperature rise gradient value, the capacitor aging index, and the connector aging index. This includes: using the product of the average temperature rise gradient value and a preset capacitor aging coefficient as the capacitor adjustment step size; using the product of the average temperature rise gradient value and a preset connector aging coefficient as the connector adjustment step size; using the difference between the preset capacitor aging index threshold and the capacitor adjustment step size as the target capacitor aging index threshold; using the difference between the preset connector aging index threshold and the connector adjustment step size as the target connector aging index threshold; determining whether the capacitor aging index is greater than the target capacitor aging index threshold; determining whether the connector aging index is greater than the target connector aging index threshold; and determining that the fault diagnosis result of the AI ​​server's power supply is an aging fault if either the capacitor aging index is greater than the target capacitor aging index threshold or the connector aging index is greater than the target connector aging index threshold. Otherwise, determining that the fault diagnosis result of the AI ​​server's power supply is no aging fault.

[0061] This specific implementation plan details how to quantitatively lower the diagnostic threshold under high-temperature deterioration environments. Different components within the power supply exhibit varying sensitivities to temperature increases. For example, electrolytic capacitors experience exponential increases in electrolyte evaporation at high temperatures, making them extremely sensitive to temperature. While connectors also experience contact problems due to thermal expansion and contraction, their sensitivity typically differs from that of capacitors. Therefore, the system first introduces preset capacitor aging coefficients and preset connector aging coefficients. These two coefficients represent the weighted proportion of increased failure risk for capacitors and connectors when facing a unit temperature increase. In practical deployment, these coefficients can be derived through accelerated aging life tests (such as the Arrhenius model) or statistical analysis of historical high-temperature failure samples. The system multiplies the average temperature rise gradient value by these two coefficients to calculate the specific penalty for different components, namely the capacitor adjustment step size and the connector adjustment step size.

[0062] Next, the system uses the calculated adjustment step size to correct the original static threshold. Specifically, the preset capacitor aging index threshold is subtracted from the capacitor adjustment step size to obtain the tightened target capacitor aging index threshold; similarly, the preset connector aging index threshold is subtracted from the connector adjustment step size to obtain the tightened target connector aging index threshold. The physical meaning of this subtraction operation is that the faster the ambient temperature deteriorates (the larger the average temperature rise gradient), the larger the adjustment step size, and the lower the final target threshold. Subsequently, the system uses the reduced target threshold to re-evaluate the current aging index. It determines whether the capacitor aging index is greater than the target capacitor aging index threshold and whether the connector aging index is greater than the target connector aging index threshold. If either index exceeds its corresponding target threshold, the system determines that the power supply has an aging fault; otherwise, it determines that there is no aging fault.

[0063] In terms of working principle, this specific implementation plan constructs a dynamic defense mechanism where "the harsher the environment, the stricter the standards." Its unique technical effect lies in achieving adaptive threshold adjustment based on the degree of environmental temperature deterioration. Under adverse operating conditions with continuously rising ambient temperatures, the system can proactively tighten health assessment standards, allowing potential hazards that are tolerable at normal temperatures but highly susceptible to sudden failures at high temperatures to be intercepted in advance. This differentiated and dynamic threshold adjustment strategy not only fully respects the differences in the thermophysical characteristics of different components but also greatly enhances the foresight of fault warnings, providing a more robust guarantee for the safe operation of the data center under extreme conditions.

[0064] like Figure 3 As shown, based on the same inventive concept as the above method, this application also proposes an AI server power supply fault diagnosis system, including: an acquisition device and a processing device; the acquisition device is used to acquire the steady-state voltage value of the AI ​​server power supply in each of multiple consecutive natural days; the processing device is used to determine the capacitor aging characteristic information of the capacitor in the power supply and the connector aging characteristic information of the connector in the power supply based on the multiple steady-state voltage values ​​of the power supply; the processing device is further used to determine the capacitor aging index of the capacitor based on the capacitor aging characteristic information of the capacitor; the processing device is further used to determine the connector aging index of the connector based on the connector aging characteristic information of the connector; the processing device is further used to determine the fault diagnosis result of the AI ​​server power supply based on the capacitor aging index and the connector aging index of the connector.

[0065] In practical applications, the aforementioned acquisition device can be implemented by the data acquisition module in the motherboard management controller (BMC), power management chip (PMIC), or rack-level intelligent power distribution unit (PDU) within the AI ​​server. This module is responsible for continuously and stably acquiring basic operational data such as voltage, current, and temperature from underlying hardware sensors. The processing device can be implemented by the AI ​​server's own central processing unit (CPU), the edge computing gateway in the local data center, or an intelligent operation and maintenance management platform deployed in the cloud. This device is responsible for executing the complex core algorithm logic, such as feature decoupling, exponential mapping, and dynamic threshold diagnosis, as described in the preceding embodiments.

[0066] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for diagnosing power supply faults in an AI server, characterized in that, include: Obtain the steady-state voltage value of the AI ​​server's power supply for each of the multiple consecutive natural days; The capacitor aging characteristics and connector aging characteristics of the power supply are determined based on multiple steady-state voltage values ​​of the power supply. The capacitor aging index is determined based on the capacitor aging characteristics. The connector aging index is determined based on the connector aging characteristics information. The fault diagnosis results of the AI ​​server's power supply are determined based on the capacitor aging index and the connector aging index.

2. The AI ​​server power supply fault diagnosis method according to claim 1, characterized in that, Obtain the steady-state voltage value of the AI ​​server's power supply for each of the multiple consecutive natural days, including: For each of the multiple natural days, obtain the original voltage value sequence and the original current value sequence of the power source within that natural day; Based on the sequence of raw current values ​​within a natural day, determine whether there is a target time period within the natural day; the target time period is a time period in which the ratio of the range of raw current values ​​to the average value of raw current values ​​within a natural day is less than a first preset ratio, and the duration is greater than or equal to the first preset duration. When a target time period exists within a natural day, the average value of all original voltage values ​​corresponding to the target time period in the original voltage value sequence is taken as the steady-state voltage value of the AI ​​server's power supply within the natural day. When there is no target time period within a natural day, obtain the device importance index of the AI ​​server, and determine the steady-state voltage value of the AI ​​server's power supply within a natural day based on the device importance index, the original voltage value sequence of the power supply within the natural day, and the original current value sequence of the power supply.

3. The AI ​​server power supply fault diagnosis method according to claim 2, characterized in that, Based on the equipment importance index, the original voltage value sequence and the original current value sequence of the power supply within a natural day, determine the steady-state voltage value of the AI ​​server's power supply within a natural day, including: The product of the equipment importance index and the preset current value coefficient is used as the initial current value step size; The target current value step size is obtained by rounding the initial current value step size up. Using the target current value step size as the step size, the original current values ​​in the original current value sequence are divided into multiple different current value intervals; Count the number of current values ​​within each current value range; The range of current values ​​with the highest number of current values ​​is taken as the target current value range; The median of the voltage values ​​corresponding to all current values ​​within the target current value range is taken as the steady-state voltage value of the AI ​​server's power supply during a natural day.

4. The AI ​​server power supply fault diagnosis method according to claim 1, characterized in that, The aging characteristics of capacitors and connectors in the power supply are determined based on multiple steady-state voltage values, including: Based on the weighted least squares method, the initial low-order polynomial curves of multiple steady-state voltage values ​​of the power supply are fitted. Determine the residuals between each steady-state voltage value of the power supply and the initial low-order polynomial curve; The steady-state voltage value with a residual greater than a preset residual threshold is taken as the target steady-state voltage value; The fitting weights of the target steady-state voltage value during the fitting of the initial low-order polynomial curve are reduced by a preset weight step size to obtain the updated fitting weights of the target steady-state voltage value. Based on the updated fitting weights of the target steady-state voltage value, the low-order polynomial curves of multiple steady-state voltage values ​​of the power supply are refitted to obtain the target low-order polynomial curves of multiple steady-state voltage values ​​of the power supply. The target low-order polynomial curve is used as the capacitor aging characteristic information. The residual sequence between the initial low-order polynomial curve and the target low-order polynomial curve is determined, and the residual sequence is used as the connector aging characteristic information of the connector.

5. The AI ​​server power supply fault diagnosis method according to claim 4, characterized in that, The capacitance aging index of a capacitor is determined based on its capacitance aging characteristics, including: Obtain the reference healthy voltage of the AI ​​server's power supply; The difference between the terminal value of the target low-order polynomial curve and the reference healthy voltage is used as the long-term cumulative offset. Perform a first-order difference operation on the target low-order polynomial curve to obtain the average slope of the target low-order polynomial curve, and use the average slope as the long-term degradation rate. Divide the long-term cumulative offset by the preset failure tolerance voltage difference of the capacitor to obtain the offset percentage; Divide the long-term degradation rate by the preset rate reference constant of the capacitor to obtain the rate hazard percentage; The weighted sum of the offset percentage and the rate hazard percentage is used as the capacitor aging index.

6. The AI ​​server power supply fault diagnosis method according to claim 4, characterized in that, The connector aging index is determined based on the connector aging characteristic information, including: Retrieve maintenance record information for each of multiple calendar days; the maintenance record information is used to indicate whether maintenance operations occurred on a given calendar day. For each calendar day, if the maintenance record information for that calendar day indicates that a maintenance operation is required, that calendar day will be used as the target calendar day. Based on linear interpolation, the new residual corresponding to the target natural day in the residual sequence is determined, and the residual corresponding to the target natural day in the residual sequence is replaced with the new residual of the target natural day to obtain the target residual sequence; The connector aging index is determined based on the target residual sequence.

7. The AI ​​server power supply fault diagnosis method according to claim 6, characterized in that, The connector aging index is determined based on the target residual sequence, including: Determine the standard deviation of the target residual sequence based on the target residual sequence; The ratio of the standard deviation of the target residual sequence to the standard deviation of the preset normal residual sequence is used as the fluctuation amplification factor; The target residual sequence is fitted with a quadratic polynomial to obtain the fitting result, and the coefficients of the quadratic term in the fitting result are extracted. The product of the fluctuation amplification factor and the preset fluctuation amplification coefficient is used as the fluctuation index. The product of the quadratic coefficient and the preset thermal runaway coefficient is used as the thermal runaway index; The weighted sum of the volatility index and the thermal runaway index is used as the connector aging index.

8. The AI ​​server power supply fault diagnosis method according to claim 1, characterized in that, The fault diagnosis results of the AI ​​server's power supply are determined based on the capacitor aging index and the connector aging index, including: Obtain the ambient temperature sequence of the AI ​​server's power supply environment for each of the multiple consecutive natural days. Determine the average temperature rise gradient value of the ambient temperature sequence; Determine whether the average temperature rise gradient value is less than the preset temperature rise gradient value; When the average temperature rise gradient value is less than the preset temperature rise gradient value, determine whether the capacitor aging index is greater than the preset capacitor aging index threshold. And determine whether the connector aging index is greater than the preset connector aging index threshold; If the capacitor aging index is greater than the preset capacitor aging index threshold, or the connector aging index is greater than the preset connector aging index threshold, the fault diagnosis result of the AI ​​server's power supply is determined to be that the power supply has an aging fault; otherwise, the fault diagnosis result of the AI ​​server's power supply is determined to be that the power supply does not have an aging fault. When the average temperature rise gradient value is greater than or equal to the preset temperature rise gradient value, the fault diagnosis result of the AI ​​server's power supply is determined based on the average temperature rise gradient value, the capacitor aging index, and the connector aging index.

9. The AI ​​server power supply fault diagnosis method according to claim 8, characterized in that, The fault diagnosis results of the AI ​​server's power supply are determined based on the average temperature rise gradient value, the capacitor aging index, and the connector aging index, including: The product of the average temperature rise gradient value and the preset capacitor aging coefficient is used as the capacitor adjustment step size, and the product of the average temperature rise gradient value and the preset connector aging coefficient is used as the connector adjustment step size. The difference between the preset capacitor aging index threshold and the capacitor adjustment step size is used as the target capacitor aging index threshold. The difference between the preset connector aging index threshold and the connector adjustment step size is used as the target connector aging index threshold. Determine whether the capacitor's aging index is greater than the target capacitor aging index threshold. Determine whether the connector aging index is greater than the target connector aging index threshold; If the capacitor aging index is greater than the target capacitor aging index threshold, or the connector aging index is greater than the target connector aging index threshold, the fault diagnosis result of the AI ​​server's power supply is determined to be an aging fault; otherwise, the fault diagnosis result of the AI ​​server's power supply is determined to be an aging fault.

10. An AI server power supply fault diagnosis system, characterized in that, include: Acquisition device and processing device; Acquisition device for acquiring the steady-state voltage value of the power supply of the AI ​​server in each of multiple consecutive natural days; The processing device is used to determine the capacitor aging characteristic information of the capacitors in the power supply and the connector aging characteristic information of the connectors in the power supply based on multiple steady-state voltage values ​​of the power supply. The processing device is also used to determine the capacitor aging index based on the capacitor aging characteristic information. The processing device is also used to determine the connector aging index of the connector based on the connector aging characteristic information of the connector; The processing device is also used to determine the fault diagnosis results of the AI ​​server's power supply based on the capacitor aging index and the connector aging index.