A server power health detection system

Abstract

The present application relates to the technical field of server power health management, and discloses a server power health detection system. The system comprises four modules connected in sequence. The health state data acquisition and synchronization module captures and standardizes power data, forming a multi-modal data stream. The fault injection and state reconstruction module drives a programmable device to inject a specific disturbance based on the data stream characteristics, and reconstructs the internal virtual operating state of the power supply according to the response. The health state evaluation and report generation module calculates the health quantitative value based on the virtual state and generates a report. The control decision and dynamic adaptation module analyzes the report, generates control instructions in combination with real-time load, and dynamically adjusts the evaluation and control parameters according to environmental changes. The system realizes early and deep diagnosis of potential faults of the power supply through active disturbance signals, and makes the health evaluation and control strategy self-adaptive to real-time operating conditions through dynamic parameter adaptation.

Description

Technical Field

[0001] This invention relates to the field of server power health management technology, specifically a server power health detection system. Background Technology

[0002] In data centers and large-scale computing systems, the reliability of server power supplies is fundamental to ensuring the stable operation of the entire IT infrastructure. Currently, the health management of server power supplies generally adopts passive monitoring technology, which involves continuously reading real-time data of key parameters such as output voltage, current, and temperature through built-in sensors and comparing them with preset fixed thresholds. Once a parameter exceeds the threshold range, the system triggers an alarm, indicating a possible fault or anomaly.

[0003] This technical approach, which relies on static thresholds and passive response data, has limitations. It struggles to effectively detect early, gradual performance degradation in internal power supply components, as these potential defects may not yet cause significant exceedances of output parameters under normal loads. Furthermore, fixed evaluation thresholds and control strategies cannot adapt to the dynamically changing and complex environment of actual server operation, easily leading to false alarms under harsh conditions or missed alarms under lenient conditions, significantly compromising the accuracy of health assessments and the rationality of control decisions.

[0004] Current technologies lack a method to proactively detect the hidden states within a power supply, and also lack a closed-loop management framework that can automatically optimize its judgments and action rules based on real-time operating context. This prevents maintenance personnel from performing accurate predictive maintenance before failures occur, and also creates bottlenecks in optimizing system energy efficiency and reliability. Summary of the Invention

[0005] The purpose of this invention is to provide a server power health detection system to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a server power supply health monitoring system, the system comprising:

[0007] The health status data acquisition and synchronization module is used to periodically capture the raw operating data of the server power supply, and perform timestamp alignment and format standardization processing on the captured raw operating data to obtain the raw multimodal data stream of the power supply.

[0008] The fault injection and state reconstruction module is used to drive the programmable fault injection device to generate a disturbance signal of a specific mode based on the inherent correlation characteristics of the original multimodal data stream of the power supply, and to reconstruct the internal virtual operating state of the power supply based on the system response after the disturbance is injected.

[0009] The health status assessment and report generation module is used to receive the internal virtual operating status of the power supply, calculate a quantitative description of the power supply health status through multi-level assessment criteria, and encapsulate and generate a power supply health status assessment report.

[0010] The control decision and dynamic adaptation module is used to parse the power health status assessment report, combine it with the preset power management strategy and real-time system load, generate power control decision commands, and dynamically adapt the assessment and control parameters according to changes in the system operating environment.

[0011] Preferably, the health status data acquisition and synchronization module obtains the raw multimodal data stream of the power supply, including:

[0012] A sensor array deployed at different monitoring locations of the server power supply captures voltage transient waveforms, current ripple characteristics, and temperature field distribution data at non-uniform intervals, forming the raw operating data.

[0013] A data synchronization coordinator is established, which assigns a unified timestamp to each set of raw operating data from the sensor array according to the network time protocol.

[0014] The data format conversion engine is invoked, which converts the original running data carrying a unified time scale into a standardized data packet with a preset data structure, sampling rate, and numerical unit.

[0015] Multiple standardized data packets are subjected to time-domain alignment and sequence splicing to generate a continuous, multi-dimensional, and time-synchronized raw multimodal data stream of the power supply.

[0016] Preferably, the fault injection and state reconstruction module reconstructs the internal virtual operating state of the power supply, including:

[0017] The programmable fault injection device is deployed in an isolated hardware sandbox environment, and the programmable fault injection device is configured with a voltage disturbance generator, a current load surge generator, and a thermal shock simulator.

[0018] The feature analysis unit parses the original multimodal data stream of the power supply and identifies the steady-state and transient characteristics of the key operating points of the power supply, which are then used as the intrinsic correlation characteristics.

[0019] Based on the inherent correlation characteristics, a disturbance mode is configured on the programmable fault injection device. The disturbance mode includes voltage drop depth, current step amplitude and temperature rise gradient, and the programmable fault injection device is driven to generate a disturbance signal of the specific mode.

[0020] The perturbation signal of the specific pattern is applied to the server power supply under test, and the power supply response signal sequence is recorded by a high-speed data acquisition card.

[0021] Based on the response signal sequence and the preset power supply physical model, the state observer algorithm is used to reconstruct a set of internal variables that cannot be directly measured. The set of internal variables constitutes the internal virtual operating state of the power supply.

[0022] Preferably, the reconstructing of the set of internal variables that cannot be directly measured using the state observer algorithm specifically includes:

[0023] The state space representation of the preset power supply physical model is constructed. This state space representation consists of system state equations and observation equations. The system state equations describe the changes of internal state variables over time, and the observation equations describe the relationship between measurable output variables and internal state variables.

[0024] Initialize the state observer, setting the initial estimates of the state variables and the observer's gain parameters;

[0025] The input and output variables in the response signal sequence are input to the state observer, which predicts the state estimate at the next moment based on the system state equation and calculates the predicted output value based on the observation equation.

[0026] Calculate the output error between the actual output value and the predicted output value, and use the output error and gain parameter to correct the state estimate to obtain the corrected state estimate.

[0027] Repeat the prediction and correction steps until the output error converges within a preset threshold range. At this point, the corrected state estimate is used as the set of internal variables that cannot be directly measured.

[0028] Preferably, the health status assessment and report generation module encapsulates the generation of a power supply health status assessment report, including:

[0029] Establish a health assessment knowledge base, which stores a set of baseline health feature vectors and a set of failure mode feature vectors corresponding to the power supply model;

[0030] Extract the current operating feature vector from the internal virtual operating state of the power supply, and input the current operating feature vector, the baseline health feature vector set, and the failure mode feature vector set into the evaluation matrix;

[0031] In the evaluation matrix, the distance measure between the current operating feature vector and the baseline health feature vector set and the failure mode feature vector set is calculated respectively;

[0032] The health status assessment engine is invoked, and the health status assessment engine outputs a quantitative description of the power supply health status based on the distance metric and using a fuzzy comprehensive evaluation method. The quantitative description of the power supply health status includes a health index and a remaining lifespan range.

[0033] The power health status quantitative description, the current operating feature vector, the timestamp for calculating the distance metric, and the corresponding evaluation matrix version number are packaged and encoded into the power health status evaluation report.

[0034] Preferably, the step of using the fuzzy comprehensive evaluation method to output a quantitative description of the power supply health status specifically includes:

[0035] A fuzzy evaluation factor set is established, which includes multiple evaluation factors derived from the distance measure, and each evaluation factor corresponds to a power health characteristic;

[0036] For each evaluation factor, a set of fuzzy levels is defined, where each fuzzy level represents a degree of fuzziness in the health status, and a membership function is assigned to each fuzzy level.

[0037] The distance metric is input into the membership function to calculate the membership value of each evaluation factor for each fuzzy level, forming a membership matrix;

[0038] Determine the weight allocation vector for the evaluation factors. The weight allocation vector is set based on expert experience or historical data and reflects the weight of each evaluation factor on the overall health status.

[0039] The membership matrix and the weight assignment vector are combined using a fuzzy synthesis operation to obtain the fuzzy evaluation result vector.

[0040] The fuzzy evaluation result vector is defuzzified, and the centroid method is used to convert the fuzzy evaluation result into an accurate health index value.

[0041] By combining the power supply's cumulative operating time and historical fault data, the remaining lifespan range is derived from the fuzzy evaluation results.

[0042] Preferably, the control decision and dynamic adaptation module generates power control decision commands, including:

[0043] Deploy a control decision model, which takes the power health status assessment report and real-time system load data as input;

[0044] The control decision model is pre-built with a decision tree and a rule engine. The decision tree is used to divide the control level according to the health index, and the rule engine is used to match specific control actions with the real-time system load data under the control level.

[0045] The control decision model executes the decision process and outputs preliminary power control decision instructions, which include output voltage fine-tuning, load switching instructions, or fan speed curves.

[0046] The system continuously acquires changes in its operating environment through an environmental awareness interface. These changes include the temperature of the computer room and the overall power consumption trend of the server cluster.

[0047] Based on changes in the system operating environment, the weight coefficients in the evaluation matrix and the rule thresholds in the control decision model are adjusted online using a parameter adaptive algorithm to complete the dynamic adaptation evaluation and control parameters, and generate the final power control decision command.

[0048] Preferably, the specific steps of deploying the control decision model, wherein the control decision model takes the power health status assessment report and real-time system load data as input, include:

[0049] Receive the power health status assessment report and parse the health index value and remaining lifespan range data from the report;

[0050] Simultaneously, real-time system load data is collected, including the current power consumption of the server rack, CPU utilization, and memory usage.

[0051] The health index value is input into the hierarchical decision tree, which has multiple health thresholds stored in the tree. The current health level of the power supply is determined by comparing the health index value with the health threshold.

[0052] Based on the health level, the corresponding rule set in the rule engine is activated. The rule set contains multiple conditional action rules, and each rule defines the control action to be executed when the system load data meets specific conditions.

[0053] The rule engine traverses the activated rule set, matches real-time system load data with rule conditions, triggers control actions of the matching rules, and generates basic control commands.

[0054] By introducing the remaining lifetime range as a correction factor, the parameters in the basic control command are scaled and adjusted to generate optimized control commands.

[0055] The optimized control command is output as a power control decision command, which is used to adjust the power supply output voltage, switch redundant power supplies, or adjust the cooling fan speed.

[0056] Preferably, the health status data acquisition and synchronization module further includes integrity verification of the raw multimodal data stream of the power supply, including:

[0057] After generating the raw multimodal data stream of the power supply, a digital digest of the raw multimodal data stream of the power supply is calculated;

[0058] The digital digest is linked and compared with the previous period's data digest pre-stored in the blockchain node;

[0059] If the comparison is consistent, the integrity of the original multimodal data stream of the power supply is confirmed, and it is marked as a valid data stream; if inconsistent, the data re-sampling mechanism is triggered to reacquire the original operating data within the corresponding time window.

[0060] Preferably, the health status assessment and report generation module further includes an online update mechanism for the assessment model, including:

[0061] The system continuously receives multiple sets of internal virtual operating states of the power supply generated after applying different disturbance modes from the fault injection and state reconstruction module.

[0062] The internal virtual operating states of the multiple power supplies are correlated and learned with the subsequent fault records of the actual power supplies.

[0063] Using the results of association learning, the failure mode feature vector set in the health assessment knowledge base is updated through incremental training, and the assessment parameters in the health status assessment engine are optimized.

[0064] Compared with the prior art, the beneficial effects of the present invention are:

[0065] Unlike conventional passive monitoring, this technology uses a programmable fault injection device to generate specific patterned disturbance signals based on the inherent correlation between power supply history and real-time data streams, applying controlled and non-destructive stimulation to the power supply. By capturing and analyzing the power supply's dynamic response to this known disturbance, its internal virtual operating state can be reconstructed. This technology enables the system to detect component characteristic drifts that are not apparent under steady-state operation, thereby achieving early and in-depth diagnosis of potential faults and elevating health management from "alarms after a fault occurs" to "early detection of performance degradation."

[0066] To address the poor adaptability of static threshold strategies in complex and changing environments, the system, while generating control decisions, possesses the ability to dynamically adjust its own assessment model parameters and control logic parameters. This adjustment is driven by parsing health assessment reports and combining them with real-time collected system load and environmental parameters. This allows the system's health judgment criteria and execution strategies to evolve in real time with contextual information such as server operating temperature, airflow, and workload. This eliminates the rigidity caused by fixed parameters, making health assessment results more closely aligned with current operating conditions and control commands more targeted. It achieves a shift from a "one-size-fits-all" fixed strategy to personalized, adaptive operation and maintenance tailored to specific environments. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the working principle of the server power health detection system described in this invention.

[0068] Figure 2 A flowchart for data acquisition and synchronization;

[0069] Figure 3 A flowchart for fault injection and state reconstruction;

[0070] Figure 4 A trend analysis chart of server power health status;

[0071] Figure 5 This is an incremental training trend diagram of the feature vector of the capacitor aging prototype. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] Please see Figure 1 This invention provides a server power supply health monitoring system. The system includes: a health status data acquisition and synchronization module that periodically captures raw operating data from multiple monitoring points of the server power supply, performs time synchronization and format unification processing on this data, and outputs a standardized raw multimodal power supply data stream. A fault injection and state reconstruction module receives this data stream, analyzes its inherent correlation characteristics, and controls a programmable fault injection device to generate a disturbance signal with a specific pattern applied to the power supply. By collecting the power supply's response signal and combining it with the power supply's physical model, the internal virtual operating state of the power supply is reconstructed. A health status assessment and report generation module receives this internal virtual operating state, calculates it using multi-layered assessment criteria, outputs a quantitative description of the power supply's health status, and encapsulates it into a structured assessment report. A control decision and dynamic adaptation module is responsible for parsing the assessment report, combining it with preset management strategies and real-time system load information, generating power supply control decision instructions, and dynamically adjusting the parameters used in the assessment and control stages according to changes in the system operating environment.

[0074] Example 1: See Figure 2The health status data acquisition and synchronization module deploys sensor arrays at various monitoring locations, including the server power supply's input, output, and key components. The sensor arrays capture voltage transient waveforms, current ripple characteristics, and temperature field distribution data at non-uniform intervals; these data constitute the raw operating data. A data synchronization coordinator within the module assigns a unified timestamp to each set of raw operating data from each sensor, based on a high-precision network time protocol. Subsequently, a data format conversion engine is invoked, converting all raw operating data with unified timestamps into standardized data packets with a preset data structure, fixed sampling rate, and standard numerical units. Multiple standardized data packets undergo time-domain alignment and sequence concatenation to generate a continuous, multi-dimensional, and time-synchronized raw multimodal power supply data stream. After generating the raw multimodal power supply data stream, the module calculates a digital digest of the data stream and links and compares this digital digest with the previous period's data digest pre-stored in the blockchain nodes. If the comparison results are consistent, the integrity of the original multimodal data stream of the power supply is confirmed, and it is marked as a valid data stream for use by subsequent modules; if the comparison results are inconsistent, the data re-acquisition mechanism is triggered, and the corresponding sensor is instructed to reacquire the original operating data within the corresponding time window.

[0075] In practical implementation, the health status data acquisition and synchronization module of the server power supply health monitoring system is implemented on a server rack configured with multiple redundant power supply units. The sensor array of the health status data acquisition and synchronization module is deployed at different key monitoring locations of each monitored power supply unit. In the specific implementation, for the server power supply with the serial number PSU-01A, a high-precision differential voltage probe is deployed at the input terminal to capture the transient waveform of the input voltage. Rogowski coils are deployed at the pins of the main power inductor and the output filter capacitor to capture the current ripple characteristics. Distributed digital temperature sensors are attached to the housing surfaces of the power switching transistors and rectifier diodes to capture temperature field distribution data. These sensors capture data according to preset non-uniform intervals: the voltage transient waveform capture period is set to 100 microseconds, the current ripple characteristic capture period is set to 200 microseconds, and the temperature field distribution data capture period is set to 1 second. The raw data streams captured in this way constitute the raw operating data for the power supply PSU-01A.

[0076] In some embodiments, the data synchronization coordinator established by the health status data acquisition and synchronization module operates under a unified timing reference. The data synchronization coordinator synchronizes to the network time protocol via a precision clock source, assigning a unified timestamp with nanosecond-level precision to each arriving frame of raw operational data. The unified timestamp format is "year-month-day hour:minute:second.nanosecond". For example, a set of transient waveform data captured by a voltage probe at a specific moment is timestamped as "2023-10-27 14:30:05.123456", while a set of readings reported by a temperature sensor at a similar time is timestamped as "2023-10-27 14:30:05.123780". It can be understood that even data from sensors from different physical locations possess temporal comparability due to the unified timestamp.

[0077] In practice, the health status data acquisition and synchronization module calls the data format conversion engine to process the raw operational data carrying a unified time stamp. The data format conversion engine has a pre-built standardized data structure template, which defines four mandatory fields: "timestamp," "data dimension," "numerical sequence," and "unit." The engine converts a string of raw analog-to-digital conversion code values ​​from the input voltage transient waveform, combined with its sensitivity coefficient, into a floating-point sequence in volts, and encapsulates it into a standardized data packet. The "unit" field of the standardized data packet is filled with "V." Similarly, the raw data of current ripple characteristics is converted into a sequence in amperes, with the "unit" field filled with "A"; temperature field distribution data is converted into a sequence in degrees Celsius, with the "unit" field filled with "°C." All standardized data packets are standardized to a sampling rate of 10,000 samples per second.

[0078] In practice, multiple standardized data packets undergo time-domain alignment and sequence splicing. The processing utilizes the "timestamp" field in the standardized data packets, constructing a virtual timeline starting with the earliest timestamp and using 1 microsecond as the basic time unit. The numerical sequence in each standardized data packet is interpolated or extracted to its corresponding position on the virtual timeline according to its timestamp, ensuring strict alignment of the voltage, current, and temperature data on the timeline. The aligned data is then spliced ​​in chronological order to generate a continuous multidimensional data matrix. This continuous multidimensional data matrix constitutes the time-synchronized raw multimodal data stream for the PSU-01A power supply. Each row of the raw multimodal data stream represents a synchronization time point, and each column represents a monitoring data type.

[0079] In practical implementation, the health status data acquisition and synchronization module performs integrity verification after generating the raw multimodal power data stream. The integrity verification process calculates a digital digest of the newly generated raw multimodal power data stream using the SHA-256 hash algorithm. The binary representation of the entire raw multimodal power data stream is taken as input, and a fixed-length 256-bit hash value is output as the digital digest for the current period. The health status data acquisition and synchronization module accesses a specific node in a permissioned blockchain network, which stores the digital digest generated in the previous verification period. The module compares the calculated current digital digest with the previous period's digital digest stored in the blockchain node. In some embodiments, if the two digital digests are completely identical, the raw multimodal power data stream generated in this period is confirmed to be complete and tamper-proof. The health status data acquisition and synchronization module marks the raw multimodal power data stream as a valid data stream and allows it to be output to the subsequent fault injection and state reconstruction module, while simultaneously adding the current digital digest as a new blockchain to the chain. Optionally, if the comparison finds that the digital digest is inconsistent, the integrity verification fails. The health status data acquisition and synchronization module immediately triggers the data re-acquisition mechanism. The data re-acquisition mechanism sends an instruction to the relevant sensor array, requiring the sensor array to recapture all the original running data within the time window from the time point corresponding to the digital digest of the previous period to the current time, and repeats the above data synchronization, format conversion, splicing and verification process using the recaptured original running data.

[0080] Example 2: See Figure 3 The fault injection and state reconstruction module deploys a programmable fault injection device in an isolated hardware sandbox environment. This device is equipped with a voltage disturbance generator, a current load surge generator, and a thermal shock simulator. The feature analysis unit within the module parses the raw multimodal data stream of the power supply, identifying the steady-state and transient characteristics of the power supply's key operating points. These characteristics are defined as the power supply's intrinsic correlation characteristics. Based on the identified intrinsic correlation characteristics, specific disturbance modes are configured on the programmable fault injection device. Disturbance mode parameters include voltage drop depth, current step amplitude, and temperature rise gradient. After configuration, the programmable fault injection device is driven to generate disturbance signals of specific modes. These disturbance signals of specific modes are applied to the server power supply under test, while a high-speed data acquisition card records the response signal sequence generated by the power supply. Based on the acquired response signal sequence and a preset power supply physical model, a state observer algorithm is used to reconstruct a set of internal variables that cannot be directly measured. This set constitutes the power supply's internal virtual operating state.

[0081] The specific implementation of the state observer algorithm includes: constructing a state-space representation of a pre-defined power supply physical model, which consists of system state equations describing the changes in internal state variables and observation equations describing the relationship between the measurable output and the internal state. Next, the state observer is initialized, setting initial estimates of the state variables and the observer's gain parameter. Then, the input and output variables from the response signal sequence are input to the state observer. The state observer predicts the state estimate for the next moment based on the system state equations and calculates the predicted output value based on the observation equations. Then, the output error between the actual output value and the predicted output value is calculated. This output error, along with the pre-defined gain parameter, is used to correct the state estimate, resulting in a corrected state estimate. The prediction and correction steps are repeated until the output error converges within a pre-defined threshold range. At this point, the corrected state estimate is used as the final reconstructed set of internal variables that cannot be directly measured.

[0082] In practice, the fault injection and state reconstruction module is implemented on a server power supply of model SRV-PS8500. The fault injection and state reconstruction module deploys a programmable fault injection device in an isolated hardware sandbox environment. The programmable fault injection device is equipped with an independent voltage disturbance generator, current load surge generator and thermal shock simulator. The voltage disturbance generator can output a rapid step or ramp disturbance with an amplitude from -50% to +20% of the rated voltage. The current load surge generator can increase the load current from 10% to 120% of the rated value in microseconds. The thermal shock simulator achieves a rapid temperature rise or drop of more than 5 degrees Celsius per minute through PAL patches attached to the surface of the power supply's key heat-generating components. The feature analysis unit receives the raw multimode power data stream for the SRV-PS8500 power supply from the health status data acquisition and synchronization module. The feature analysis unit analyzes the raw multimode power data stream for ten consecutive minutes and identifies that the ripple frequency characteristic of the DC voltage output by the power supply during steady-state operation is 100 kHz. It also identifies that the current spike amplitude generated by the main power transistor during switching is 1.5 times the rated current and that the surface temperature gradient distribution of the heat sink shows a pattern of high at the center and low at the edges. These steady-state and transient characteristics are defined as the intrinsic correlation characteristics of the raw multimode power data stream. Based on the identified inherent correlation characteristics, a set of disturbance modes is configured on the programmable fault injection device. The disturbance mode sets the voltage drop depth to 30% of the rated output voltage and the duration to 10 milliseconds. The disturbance mode sets the current step amplitude to a step from 50% load to 100% load and the rise time to 100 microseconds. The disturbance mode sets the temperature rise gradient to 8 degrees Celsius per minute, starting from a 60-degree Celsius reference point. After configuration, the programmable fault injection device is driven to generate a disturbance signal containing the above-mentioned voltage drop, current step, and temperature rise gradient composite specific mode.

[0083] In some embodiments, a specific pattern of disturbance signal is applied to the SRV-PS8500 server power supply under test. A voltage disturbance generator of a programmable fault injection device is connected in parallel to the DC output of the power supply, and a current load surge converter is connected in series in the load circuit. The Peltier patch of the thermal shock simulator is attached to the power MOSFET housing. Simultaneously with the disturbance application, a 16-bit high-speed data acquisition card with a sampling rate of 1 MHz synchronously records the voltage response signal sequence, current response signal sequence, and temperature sensor feedback signal sequence at the power supply output. All response signal sequences are accompanied by a high-precision timestamp synchronized with the disturbance signal. In a specific implementation, based on the recorded response signal sequences and a preset SRV-PS8500 power supply physical model, a Luneburg state observer algorithm is used to reconstruct the set of internal variables that cannot be directly measured. The preset power supply physical model is a nonlinear model describing the topology of a buck converter. A state-space representation of a pre-defined power supply physical model is constructed. The state-space representation consists of system state equations and observation equations. The system state equations describe the changes of internal state variables such as inductor current and output capacitor voltage over time. The observation equations describe the relationship between the directly measurable output voltage and total input current and these internal state variables.

[0084] Initialize the Luneburg state observer, setting the initial estimate of the inductor current state variable to zero and the initial estimate of the output capacitor voltage state variable to the rated output voltage. Set the observer's gain parameter matrix to a constant matrix determined based on the model's pole configuration. Input and output variables from the response signal sequence are fed into the Luneburg state observer. The observer predicts the state estimate for the next moment based on the system state equation and calculates the predicted output voltage and input current values ​​based on the observation equation. Calculate the output error between the actual acquired output voltage and input current values ​​and the predicted output values. Use the output error and the gain parameter matrix to correct the state estimates, obtaining the corrected inductor current and capacitor voltage state estimates. Repeat the prediction and correction steps until the Euclidean norm of the output error converges to a preset threshold range of 0.01. At this point, the corrected set of state estimates, including the estimated waveform of the switching node voltage, the estimated value of the power semiconductor junction temperature, and the estimated value of the equivalent series resistance of the filter capacitor, is used as a reconstructed set of internal variables that cannot be directly measured. This set of internal variables constitutes the internal virtual operating state of the SRV-PS8500 power supply. Understandably, the dynamic correction process of the Lundberg state observer enables the internal state estimation to track the behavior of the real system.

[0085] The core correction step of the Luneburg state observer algorithm is described by the following formula:

[0086]

[0087] Where: symbol Represents at discrete time points Corrected state estimation vector, sign Represented by time point Timing of Information Prediction State estimation vector, symbol Represents the observer gain matrix, symbol Represents a point in time The actual measured output vector, sign Represents the observation matrix. State estimation vector. Includes inductor current estimates and capacitor voltage estimates, output vector Includes measured output voltage and input current values. Optional, observer gain matrix. By employing the pole placement method, the eigenvalues ​​of the dynamic system with state estimation error are positioned at specific locations in the left half of the complex plane. It can be understood that through iterative calculations using this formula, the state estimation vector... Gradually approximate the actual internal state variables of the power supply.

[0088] Optionally, during the reconstruction of the power supply's internal virtual operating state, if the response signal sequence shows that the output voltage overshoot exceeds 5% under a current step disturbance, the state observer algorithm will adjust the estimated values ​​of the filter capacitor parameters in its internal model accordingly, thereby more accurately reflecting the impact of capacitor aging in subsequent prediction steps. The fault injection and state reconstruction module ultimately outputs a data file containing a time series of internal variables, which represents the reconstructed internal virtual operating state of the power supply under a specific disturbance test.

[0089] Example 3: The health status assessment and report generation module establishes a health assessment knowledge base, which stores a baseline health feature vector set and a failure mode feature vector set corresponding to the current server power supply model. The module extracts the current operating feature vector from the power supply's internal virtual operating state and inputs the current operating feature vector, the baseline health feature vector set, and the failure mode feature vector set into the assessment matrix. In the assessment matrix, the distance measure between the current operating feature vector and the baseline health feature vector set and the failure mode feature vector set is calculated. Subsequently, the health status assessment engine is invoked. Based on the calculated distance measure, the health status assessment engine uses a fuzzy comprehensive evaluation method to output a quantitative description of the power supply's health status, which includes a health index and remaining lifespan range.

[0090] The specific process of applying the fuzzy comprehensive evaluation method includes: establishing a fuzzy evaluation factor set, which contains multiple evaluation factors derived from distance measures, each corresponding to a power supply health characteristic. For each evaluation factor, a fuzzy level set is defined, where each fuzzy level represents a degree of fuzziness in the health status, and a membership function is assigned to each fuzzy level. The distance measure is input into each membership function, and the membership value of each evaluation factor for each fuzzy level is calculated to form a membership matrix. The weight allocation vector of the evaluation factors is determined, which is set based on expert experience or historical data, reflecting the influence weight of each evaluation factor on the overall health status. The membership matrix and the weight allocation vector are subjected to fuzzy synthesis operation to obtain a fuzzy evaluation result vector. The fuzzy evaluation result vector is defuzzified, and the centroid method is used to convert the fuzzy evaluation result into a precise health index value. Combining the power supply's cumulative operating time and historical fault data, the remaining lifespan range is derived from the fuzzy evaluation result. The module packages and encodes the quantitative description of the power supply health status, the current operating characteristic vector, the timestamp when calculating the distance measure, and the corresponding evaluation matrix version number into a power supply health status evaluation report.

[0091] In practical implementation, the health status assessment and report generation module processes the internal virtual operating status of the SRV-PS8500 server power supply output by the fault injection and status reconstruction module. The health status assessment and report generation module has an internal health assessment knowledge base, which stores a baseline health feature vector set and a failure mode feature vector set that completely correspond to the SRV-PS8500 power supply model. The baseline health feature vector set contains multiple feature vectors extracted from typical internal variable data collected from the power supply in its brand-new factory condition. For example, there is a feature vector representing the output ripple under ideal conditions [ripple voltage peak 10mV, ripple frequency 100kHz], and a feature vector representing ideal efficiency [full load efficiency 95.2%, half load efficiency 96.1%]. The failure mode feature vector set contains feature vectors summarized from historical failure cases. For example, there is a feature vector representing the failure of the output filter capacitor due to drying out [the equivalent series resistance of the capacitor increases to 200mΩ, and the peak ripple voltage increases to 150mV], and a feature vector representing the increase in on-resistance due to aging of the power MOSFET [switching losses increase by 15%, and the temperature rise rate increases by 30%].

[0092] In practical implementation, a set of current operating feature vectors is extracted from the internal virtual operating state of the SRV-PS8500 power supply. These current operating feature vectors include the peak ripple voltage of 12.5mV calculated from the reconstructed switching node voltage waveform, the current full-load efficiency of 94.8% obtained from the efficiency calculation module, the temperature rise rate of 4.5 degrees Celsius per minute calculated from the temperature rise curve, the load step recovery time of 50 microseconds extracted from the dynamic response data, and the estimated equivalent series resistance of 80mΩ from the state observer output. The current operating feature vector [12.5mV, 94.8%, 4.5°C / min, 50μs, 80mΩ], the corresponding vector from the baseline health feature vector set, and the relevant vector from the failure mode feature vector set are input together into the evaluation matrix. In the evaluation matrix, the distance measure between the current operating feature vector and the baseline health feature vector set and the failure mode feature vector set is calculated. The distance measure is calculated using weighted Euclidean distance. For example, the distance between the current ripple voltage peak of 12.5mV and the baseline feature of 10mV, and the distance between the current ripple voltage peak of 12.5mV and the capacitor failure feature of 150mV are calculated.

[0093] The health status assessment engine is invoked. Based on multiple calculated distance measures, the engine uses a fuzzy comprehensive evaluation method to output a quantitative description of the power supply's health status. The specific process of using the fuzzy comprehensive evaluation method includes: first, establishing a fuzzy evaluation factor set, which contains five evaluation factors derived from the distance measures: output ripple deviation, efficiency degradation, temperature rise anomaly, response deterioration, and component aging. For each evaluation factor, a fuzzy level set is defined, consisting of four levels: "Health," "Attention," "Warning," and "Fault," and a trapezoidal membership function is assigned to each fuzzy level. The calculated distance measures are input into the membership function corresponding to each evaluation factor. For example, inputting the distance value of the output ripple deviation results in a membership degree of 0.7 for the "Health" level, 0.3 for the "Attention" level, and 0 for both the "Warning" and "Fault" levels, forming a 5x4 membership degree matrix.

[0094] A weight allocation vector for the evaluation factors was determined. Based on the design expertise of SRV-PS8500 power supplies and extensive historical operating data, the weight allocation vector was set to [0.25, 0.20, 0.25, 0.15, 0.15]. This weight allocation vector reflects that output ripple deviation and temperature rise anomaly have the greatest impact on the overall health status, followed by efficiency degradation, and then response deterioration and component aging. A fuzzy synthesis operation was performed between the membership matrix and the weight allocation vector using a weighted average operator to obtain the fuzzy evaluation result vector [0.62, 0.28, 0.10, 0.00]. The fuzzy evaluation result vector was then defuzzified, and the centroid method was used to convert the fuzzy evaluation result into a precise health index value, which was 82. Combining the cumulative operating time (12000 hours) and historical fault-free records of the SRV-PS8500 power supply, the remaining lifespan range of [8000, 12000] hours was derived from the fuzzy evaluation results.

[0095] The centroid method used in defuzzification is described by the following formula:

[0096]

[0097] Where: symbol Represents the calculated health index value, symbol The total number of fuzzy levels, in this implementation ,symbol The first digit in the fuzzy evaluation result vector represents the ... The membership degree value corresponding to each fuzzy level, symbol Representative for the first A representative value is set for a fuzzy level, for example, a representative value for the "healthy" level. "Attention" level represents a value "Warning" level representative value "Fault" level represents the value .

[0098] The calculated power supply health status quantification description (health index 82, remaining lifespan range [8000, 12000] hours), current operating feature vector [12.5mV, 94.8%, 4.5°C / min, 50μs, 80mΩ], timestamp of distance measurement calculation "2023-10-27 15:22:31", and evaluation matrix version number "V2.1.5" used in this evaluation are packaged and encoded into a structured JSON format data packet. This JSON format data packet is the power supply health status evaluation report for this SRV-PS8500 power supply. It can be understood that the power supply health status evaluation report contains complete information such as quantitative conclusions, original characteristics, evaluation time, and model version. Optionally, when the distance measure of a certain evaluation factor, such as "component aging degree," is very close to the failure mode feature vector, its membership degree in the "failure" level will significantly increase. This will cause the membership degree of the "failure" level in the final fuzzy evaluation result vector to no longer be zero, and consequently, the health index value calculated by the centroid method will decrease accordingly, and the remaining lifespan range will also be shortened. In some embodiments, the failure mode feature vector set stored in the health assessment knowledge base contains more than one failure type. The evaluation matrix calculates the distance between the current operating feature vector and each failure mode feature vector, and selects the smallest distance value to participate in the subsequent membership degree calculation. This allows the evaluation to point to the most likely specific failure mode. Optionally, the weight allocation vector is not fixed, but in this specific implementation, the weight allocation vector is valued according to a preset fixed strategy.

[0099] See Figure 4 This is a chart analyzing the health status of a server power supply. Its core content shows the correlation between power supply uptime and health indicators. The initial health index is close to 98, gradually decreasing with uptime, and stabilizing in the 81-82 range after 12,000 hours. The initial temperature rise rate is approximately 8.1°C / min, gradually increasing to approximately 4.3°C / min after 12,000 hours. The initial peak ripple voltage is approximately 92.5mV, continuously rising to approximately 96mV after 12,000 hours. This chart belongs to the category of equipment health monitoring charts and is commonly used in the operation and maintenance management of server power supplies. It illustrates the aging pattern of "the longer the uptime, the lower the health index and the worse the temperature rise / ripple," which can help determine the remaining lifespan of the power supply and provide early warnings of potential failures.

[0100] Example 4: The control decision and dynamic adaptation module deploys a control decision model that takes a power health status assessment report and real-time system load data as input. The control decision model has a pre-built decision tree and rule engine. The decision tree is used to classify control levels based on health indices, and the rule engine is used to match specific control actions based on real-time system load data at a given control level. The control decision model executes the decision process and outputs preliminary power control decision commands, which may include output voltage fine-tuning, load switching commands, or fan speed curves. The module continuously acquires information on changes in the system operating environment through an environmental awareness interface, such as the ambient temperature of the computer room and the overall power consumption trend of the server cluster. Based on these changes in the system operating environment, the module uses a parameter adaptive algorithm to adjust the weight coefficients in the evaluation matrix and the rule thresholds in the control decision model online, completing the dynamic adaptation assessment and control parameters, and generating the final power control decision commands. The specific steps of the control decision model, taking a power health status assessment report and real-time system load data as input, include receiving the power health status assessment report and parsing the health index value and remaining lifespan interval data from the report. Simultaneously, it collects real-time system load data, including the current power consumption of the server rack, CPU utilization, and memory usage.

[0101] The health index value is input into a hierarchical decision tree, which pre-stores multiple health thresholds. By comparing the health index value with these thresholds, the current health level of the power supply is determined. Based on the determined health level, the corresponding rule set in the rule engine is activated. The rule set contains multiple conditional action rules, each defining the control action to be performed when the system load data meets specific conditions. The rule engine traverses the activated rule set, matching the real-time system load data with the rule conditions, triggering the control action of the matching rule, and generating basic control commands. The remaining lifetime range is introduced as a correction factor to scale and adjust the parameters in the basic control commands, generating optimized control commands. The optimized control commands are output as the final power supply control decision commands, used to adjust the power supply's output voltage, switch redundant power supplies, or adjust the cooling fan speed.

[0102] In practice, the control decision and dynamic adaptation module is implemented for an SRV-PS8500 server power supply with a health index of 82 and a remaining lifespan of [8000, 12000] hours, and its rack. The module deploys a control decision model, which takes the health index value of 82 (parsed from the power supply health status assessment report), the remaining lifespan of [8000, 12000] hours, and real-time system load data collected from the server management bus as input. The real-time system load data includes the current total power consumption of the server rack (1250 watts), CPU utilization of 65%, and memory utilization of 70%. The control decision and dynamic adaptation module executes the specific steps of receiving the power supply health status assessment report and parsing the health index value and remaining lifespan data. The control decision model inputs the health index value of 82 into its internally pre-set hierarchical decision tree, which stores health thresholds of 90, 75, and 60. The hierarchical decision tree compares the health index value of 82 with the health threshold, determining that the health index value of 82 is greater than 75 but less than 90, and therefore classifies the current health level of the power supply as "Attention". Based on the "Attention" health level, the control decision model activates the rule set corresponding to the "Attention" level in the rule engine, which contains multiple conditional action rules.

[0103] The rule engine traverses the activated rule set, matching real-time system load data with rule conditions. One rule condition is "CPU utilization greater than 60% and memory utilization greater than 65%." The current CPU utilization of 65% and memory utilization of 70% meet this condition, triggering the control action of "increasing fan speed by one level," thus generating the basic control instruction "increase fan speed by one level." A remaining lifespan range of [8000, 12000] hours is introduced as a correction factor. This remaining lifespan range is in the lower-middle range. The correction factor scales and adjusts the parameters in the basic control instruction, increasing the specific increment of "increase fan speed by one level" from the preset 200 RPM to 300 RPM, generating the optimized control instruction "increase fan speed by 300 RPM." The control decision model outputs this optimized control instruction as the power control decision instruction, which is sent to the fan controller of the SRV-PS8500 power supply through the power management interface. Refer to Table 1 for the correspondence between health levels and potential control actions in the hierarchical decision tree.

[0104] Table 1: Mapping Table between Health Levels and Control Strategies

[0105]

[0106] In practical implementation, the pre-built decision tree and rule engine within the control decision model work together. The decision tree is used to classify control levels based on health indices, while the rule engine is used to match specific control actions based on real-time system load data at each control level. The control decision model executes the decision-making process, outputting preliminary power control decision commands, which may include output voltage fine-tuning, load switching commands, or fan speed curves. The control decision and dynamic adaptation module continuously acquires information about changes in the system operating environment through an environmental awareness interface. This interface reads the current data center temperature as 28 degrees Celsius from the data center monitoring system and the overall power consumption trend of the server cluster from the cluster management platform, showing an upward trend over the past hour. Based on these changes in the system operating environment, the control decision and dynamic adaptation module uses a parameter adaptive algorithm to adjust the weight coefficients in the evaluation matrix and the rule thresholds in the control decision model online. It is understandable that increased data center temperature exacerbates the power supply's heat dissipation pressure.

[0107] The parameter adaptive algorithm dynamically adjusts parameters according to changes in the system operating environment. A specific formula for its online adjustment rule threshold is described as follows:

[0108]

[0109] Where: symbol Represents the adjusted new rule threshold, symbol A baseline value representing a rule threshold, such as the CPU utilization baseline threshold that triggers fan acceleration. ,symbol Represents the influence coefficient of ambient temperature (e.g.) ),symbol This represents the currently perceived ambient temperature of the computer room (28°C), symbol... This represents the reference design temperature for the data center environment (25°C). Substituting the current ambient temperature of 28 degrees Celsius, the new rule threshold is calculated. After completing the dynamic adaptation evaluation and control parameters, the control decision model re-matches the rules based on the new rule thresholds and generates the final fan speed control command.

[0110] Example 5: The health status assessment and report generation module includes an online update mechanism for the assessment model. This mechanism continuously receives multiple sets of internal virtual operating states of power supplies generated after applying different perturbation modes from the fault injection and state reconstruction module. These multiple sets of internal virtual operating states of power supplies are then correlated with subsequent actual fault records of the server power supplies. Using the results of this correlation learning, the failure mode feature vector set in the health assessment knowledge base is updated through incremental training, and the assessment parameters in the health status assessment engine are optimized.

[0111] In practical implementation, the online update mechanism of the health status assessment and report generation module operates on a server power supply cluster of model SRV-PS8500. This mechanism continuously receives multiple sets of internal virtual operating states of the power supplies from the fault injection and status reconstruction module. These internal virtual operating states are derived from test results after applying different perturbation modes to multiple SRV-PS8500 power supplies of the same model. For example, the first set of received internal virtual operating states corresponds to the reconstructed data after injecting a 30% voltage drop perturbation; the second set corresponds to the reconstructed data after injecting a 50% step load current perturbation; and the third set corresponds to the reconstructed data after injecting a temperature rise gradient of 8 degrees Celsius per minute perturbation. The online update mechanism of the assessment model correlates these multiple sets of internal virtual operating states of the power supplies with subsequent fault records generated by the corresponding power supplies in the actual operating environment. These fault records are derived from the server maintenance logs, which record an event where a power supply experienced output voltage instability after 15,000 hours of operation, ultimately confirmed as a failure of the main filter capacitor.

[0112] In practice, the association learning process compares and analyzes the "main filter capacitor failure" event recorded in the fault log with the internal virtual operating states of multiple power supplies previously received. Association learning identifies that in several monitoring sessions prior to the fault, the parameter representing the equivalent series resistance of the capacitor in the internal virtual operating state of that power supply showed a continuous increasing trend, and the harmonic component characteristics of its output ripple voltage were highly similar to those observed in the internal virtual operating state of the second power supply after the injected voltage drop disturbance. Using the results of association learning, the online update mechanism of the evaluation model updates the failure mode feature vector set in the health assessment knowledge base through incremental training. Incremental training supplements or corrects the original feature vectors based on newly identified fault mode features. For example, a new feature vector for "early capacitor aging" is added to the failure mode feature vector set. This feature vector includes two key dimensions: "monthly growth rate of equivalent series resistance of capacitor exceeds 5%" and "third harmonic ripple amplitude accounts for more than 15% of the fundamental frequency." Meanwhile, the online update mechanism of the evaluation model optimizes the evaluation parameters in the health status evaluation engine, especially by adjusting the membership function parameters related to the "component aging" evaluation factor in the fuzzy comprehensive evaluation method. This allows the model to be mapped to the fuzzy level of "early warning" or "fault" more quickly when a similar capacitor equivalent series resistance growth pattern and harmonic characteristics are detected again.

[0113] The online update mechanism of the evaluation model employs an incremental training approach that adjusts the prototype vector of the failure mode feature vector set through a recursive formula, which is described as follows:

[0114]

[0115] Where: symbol This indicates that the updated category belongs to... The prototype feature vector, symbol This represents the prototype feature vector of this category stored in the health assessment knowledge base before the update, with the symbol... The learning rate is a constant between 0 and 1, with the sign... This represents the new feature vector observed and extracted from the internal virtual operating state of the power supply in the newly associated fault case. Learning rate. It controls the speed at which new knowledge is incorporated. This formula allows the prototype vectors in the failure mode feature vector set to progressively adapt to newly emerging failure mode features without requiring a complete retraining of the model.

[0116] See Figure 5 This is an incremental training trend chart of the prototype feature vector of capacitor aging, used to show the update process of model parameters under different learning rates. The first two iterations quickly increased from 50 to 70, and it nearly converged after about 6 iterations; the update speed was moderate, reaching about 73 after 10 iterations; the update speed was the slowest, reaching only about 69.5 after 10 iterations. A larger learning rate results in a higher final converged feature vector value and an earlier plateau. This chart is an effect analysis chart of incremental model training, applied to the "evaluation model online update" stage of a server power health monitoring system. The learning rate is a key parameter in incremental training, controlling the speed at which new fault features are incorporated into the model; the chart can help select an appropriate learning rate; it reflects the optimization process of the prototype feature vector of the "capacitor aging" failure mode with training iterations, ultimately enabling the model to more accurately identify early capacitor aging faults.

[0117] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A server power supply health monitoring system, characterized in that, The system includes: The health status data acquisition and synchronization module is used to periodically capture the raw operating data of the server power supply, and perform timestamp alignment and format standardization processing on the captured raw operating data to obtain the raw multimodal data stream of the power supply. The fault injection and state reconstruction module is used to drive the programmable fault injection device to generate a disturbance signal of a specific mode based on the inherent correlation characteristics of the original multimodal data stream of the power supply, and to reconstruct the internal virtual operating state of the power supply based on the system response after the disturbance is injected. The health status assessment and report generation module is used to receive the internal virtual operating status of the power supply, calculate a quantitative description of the power supply health status through multi-level assessment criteria, and encapsulate and generate a power supply health status assessment report. The control decision and dynamic adaptation module is used to parse the power health status assessment report, combine the preset power management strategy and real-time system load to generate power control decision instructions, and dynamically adapt the assessment and control parameters according to changes in the system operating environment. The fault injection and state reconstruction module reconstructs the internal virtual operating state of the power supply, including: The programmable fault injection device is deployed in an isolated hardware sandbox environment, and the programmable fault injection device is configured with a voltage disturbance generator, a current load surge generator, and a thermal shock simulator. The feature analysis unit parses the original multimodal data stream of the power supply and identifies the steady-state and transient characteristics of the key operating points of the power supply, which are then used as the intrinsic correlation characteristics. Based on the inherent correlation characteristics, a disturbance mode is configured on the programmable fault injection device. The disturbance mode includes voltage drop depth, current step amplitude and temperature rise gradient, and the programmable fault injection device is driven to generate a disturbance signal of the specific mode. The perturbation signal of the specific pattern is applied to the server power supply under test, and the power supply response signal sequence is recorded by a high-speed data acquisition card. Based on the response signal sequence and the preset power supply physical model, the state observer algorithm is used to reconstruct a set of internal variables that cannot be directly measured. The set of internal variables constitutes the internal virtual operating state of the power supply.

2. The server power supply health detection system as described in claim 1, characterized in that, The health status data acquisition and synchronization module obtains the raw multimodal data stream of the power supply, including: A sensor array deployed at different monitoring locations of the server power supply captures voltage transient waveforms, current ripple characteristics, and temperature field distribution data at non-uniform intervals, forming the raw operating data. A data synchronization coordinator is established, which assigns a unified timestamp to each set of raw operating data from the sensor array according to the network time protocol. The data format conversion engine is invoked, which converts the original running data carrying a unified time scale into a standardized data packet with a preset data structure, sampling rate, and numerical unit. Multiple standardized data packets are subjected to time-domain alignment and sequence splicing to generate a continuous, multi-dimensional, and time-synchronized raw multimodal data stream of the power supply.

3. The server power supply health detection system as described in claim 2, characterized in that, The process of reconstructing the set of internal variables that cannot be directly measured using the state observer algorithm specifically includes: The state space representation of the preset power supply physical model is constructed. This state space representation consists of system state equations and observation equations. The system state equations describe the changes of internal state variables over time, and the observation equations describe the relationship between measurable output variables and internal state variables. Initialize the state observer, setting the initial estimates of the state variables and the observer's gain parameters; The input and output variables in the response signal sequence are input to the state observer, which predicts the state estimate at the next moment based on the system state equation and calculates the predicted output value based on the observation equation. Calculate the output error between the actual output value and the predicted output value, and use the output error and gain parameter to correct the state estimate to obtain the corrected state estimate. Repeat the prediction and correction steps until the output error converges within a preset threshold range. At this point, the corrected state estimate is used as the set of internal variables that cannot be directly measured.

4. The server power supply health detection system as described in claim 3, characterized in that, The health status assessment and report generation module encapsulates and generates a power supply health status assessment report, including: Establish a health assessment knowledge base, which stores a set of baseline health feature vectors and a set of failure mode feature vectors corresponding to the power supply model; Extract the current operating feature vector from the internal virtual operating state of the power supply, and input the current operating feature vector, the baseline health feature vector set, and the failure mode feature vector set into the evaluation matrix; In the evaluation matrix, the distance measure between the current operating feature vector and the baseline health feature vector set and the failure mode feature vector set is calculated respectively; The health status assessment engine is invoked, and the health status assessment engine outputs a quantitative description of the power supply health status based on the distance metric and using a fuzzy comprehensive evaluation method. The quantitative description of the power supply health status includes a health index and a remaining lifespan range. The power health status quantitative description, the current operating feature vector, the timestamp for calculating the distance metric, and the corresponding evaluation matrix version number are packaged and encoded into the power health status evaluation report.

5. A server power supply health detection system as described in claim 4, characterized in that, The step of using the fuzzy comprehensive evaluation method to output a quantitative description of the power supply health status specifically includes: A fuzzy evaluation factor set is established, which includes multiple evaluation factors derived from the distance measure, and each evaluation factor corresponds to a power health characteristic; For each evaluation factor, a set of fuzzy levels is defined, where each fuzzy level represents a degree of fuzziness in the health status, and a membership function is assigned to each fuzzy level. The distance metric is input into the membership function to calculate the membership value of each evaluation factor for each fuzzy level, forming a membership matrix; Determine the weight allocation vector for the evaluation factors. The weight allocation vector is set based on expert experience or historical data and reflects the weight of each evaluation factor on the overall health status. The membership matrix and the weight assignment vector are combined using a fuzzy synthesis operation to obtain the fuzzy evaluation result vector. The fuzzy evaluation result vector is defuzzified, and the centroid method is used to convert the fuzzy evaluation result into an accurate health index value. By combining the power supply's cumulative operating time and historical fault data, the remaining lifespan range is derived from the fuzzy evaluation results.

6. The server power supply health detection system as described in claim 5, characterized in that, The control decision and dynamic adaptation module generates power control decision commands, including: Deploy a control decision model, which takes the power health status assessment report and real-time system load data as input; The control decision model is pre-built with a decision tree and a rule engine. The decision tree is used to divide the control level according to the health index, and the rule engine is used to match specific control actions with the real-time system load data under the control level. The control decision model executes the decision process and outputs preliminary power control decision instructions, which include output voltage fine-tuning, load switching instructions, or fan speed curves. The system continuously acquires changes in its operating environment through an environmental awareness interface. These changes include the temperature of the computer room and the overall power consumption trend of the server cluster. Based on changes in the system operating environment, the weight coefficients in the evaluation matrix and the rule thresholds in the control decision model are adjusted online using a parameter adaptive algorithm to complete the dynamic adaptation evaluation and control parameters, and generate the final power control decision command.

7. A server power supply health detection system as described in claim 6, characterized in that, The specific steps for deploying the control decision model, which uses the power health status assessment report and real-time system load data as input, include: Receive the power health status assessment report and parse the health index value and remaining lifespan range data from the report; Simultaneously, real-time system load data is collected, including the current power consumption of the server rack, CPU utilization, and memory usage. The health index value is input into the hierarchical decision tree, which has multiple health thresholds stored in the tree. The current health level of the power supply is determined by comparing the health index value with the health threshold. Based on the health level, the corresponding rule set in the rule engine is activated. The rule set contains multiple conditional action rules, and each rule defines the control action to be executed when the system load data meets specific conditions. The rule engine traverses the activated rule set, matches real-time system load data with rule conditions, triggers control actions of the matching rules, and generates basic control commands. By introducing the remaining lifetime range as a correction factor, the parameters in the basic control command are scaled and adjusted to generate optimized control commands. The optimized control command is output as a power control decision command, which is used to adjust the power supply output voltage, switch redundant power supplies, or adjust the cooling fan speed.

8. The server power supply health detection system as described in claim 7, characterized in that, The health status data acquisition and synchronization module also includes integrity verification of the raw multimodal data stream of the power supply, including: After generating the raw multimodal data stream of the power supply, a digital digest of the raw multimodal data stream of the power supply is calculated; The digital digest is linked and compared with the previous period's data digest pre-stored in the blockchain node; If the comparison is consistent, the integrity of the original multimodal data stream of the power supply is confirmed, and it is marked as a valid data stream; if inconsistent, the data re-sampling mechanism is triggered to reacquire the original operating data within the corresponding time window.

9. A server power supply health detection system as described in claim 8, characterized in that, The health status assessment and report generation module also includes an online update mechanism for the assessment model, including: The system continuously receives multiple sets of internal virtual operating states of the power supply generated after applying different disturbance modes from the fault injection and state reconstruction module. The internal virtual operating states of the multiple power supplies are correlated and learned with the subsequent fault records of the actual power supplies. Using the results of association learning, the failure mode feature vector set in the health assessment knowledge base is updated through incremental training, and the assessment parameters in the health status assessment engine are optimized.

Images