Fault self-diagnosis modular direct-current power supply system

Abstract

The invention relates to the technical field of direct-current power supply fault diagnosis, in particular to a fault self-diagnosis modular direct-current power supply system, which is characterized in that a failure mechanism dynamic modeling unit constructs a multi-stress coupling degradation mapping library, adopts a degradation mechanism decoupling algorithm to separate and superpose failure effects and generates a dynamic degradation model; the early degradation capture unit deploys a high-frequency wide-domain sampling circuit and a noise suppression signal chain, combines time-frequency domain joint feature extraction and separates degradation feature signals from background noise, and the self-adaptive life prediction unit adopts a long short-term memory network and random forest fusion algorithm to dynamically correct parameters and output residual life probability distribution. The intelligent fault judgment unit realizes progressive diagnosis through a hierarchical threshold decision mechanism, and the health management cooperative execution unit triggers a pre-maintenance strategy, executes power flexible derating, redundancy switching or directional fusing isolation, and improves the reliability of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of direct current power supply fault diagnosis, in particular to a modular direct current power supply system with self-diagnosis function. BACKGROUND

[0002] Direct current power supply fault diagnosis is an important technology. Under the background of the current new power system construction promotion and the significant improvement of power grid demand for power supply reliability and intelligent operation and maintenance, this technology is the key support to break through the limitations of traditional direct current power supply. It can not only avoid the shutdown of power grid equipment and the interruption of production caused by power supply failure, but also reduce resource waste through fine diagnosis and preventive maintenance. At the same time, it meets the requirements of equipment active safety and intelligent operation and maintenance technology upgrading in the State Grid "Smart AC / DC Power Supply System Construction Scheme", and provides core technical support for the digital transformation of direct current power supply system.

[0003] Under normal circumstances, alternating current power supply is rectified by high-frequency switching power supply to direct current power supply to supply power to various types of equipment, and at the same time, the battery pack is charged. When the alternating current power supply is interrupted, the battery pack immediately works to ensure the continuous operation of the key equipment and guarantee the safe and stable operation of the system. Its typical structure diagram is shown in Figure 3

[0004] ​The existing modular DC power supply system faces the problems of insufficient reliability of charging devices, lagging fault diagnosis, and limited DC feeder maintenance in actual application, which is difficult to meet the core problems of high reliability and intelligent operation and maintenance requirements. Among them, the charging device as the core power supply unit has poor adaptability to a wide range of input voltage, low hardware design redundancy, and is prone to power supply instability under power grid fluctuations or extreme working conditions, insufficient parallel current sharing accuracy and adaptability, relies on a single precision detection technology when multiple modules are connected in parallel, has low resolution under small current conditions and is prone to over-difference under large current conditions, and lacks adaptive strategies for software and hardware collaboration, resulting in reduced system efficiency and shortened module life. The detection of core devices such as power tubes only stays at the module level, lacks accurate diagnosis technology at the device level, and early degradation signals are easily masked by noise due to their weak amplitude, making it difficult to capture in a timely manner and causing safety hazards such as thermal imbalance. At the same time, the DC feeder screen as the core of power distribution also has operation and maintenance pain points, that is, maintenance requires power outage operation, the feeder switch, sensor and main circuit are strongly coupled, and replacing components requires disconnecting the load, affecting power supply continuity. The secondary line process is complex and has a high error rate, the load transfer and safety protection work are greatly reduced when operating under voltage, the information perception dimension is single, only the on-off state and total leakage current are collected, and there is a lack of deep data such as branch-level load current and temperature, which makes it difficult to support preventive maintenance. These problems are layered and stacked, which not only easily causes power grid equipment to shut down due to sudden failures, but also may cause component waste due to excessive reliance on fixed maintenance cycles, which not only meets the requirements of new power systems for high reliability, but also meets the technical upgrade requirements of State Grid smart operation and maintenance. In order to solve this technical problem, we provide a fault self-diagnosis modular DC power supply system. SUMMARY

[0005] The purpose of the present application is to provide a fault self-diagnosis modular DC power supply system to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, a fault self-diagnosis modular DC power supply system is provided, comprising: The failure mechanism dynamic modeling unit associates different environmental stresses and working conditions to the failure physical model of a specific component by constructing a multi-stress coupling degradation mapping library, separates and superimposes the failure effects using a degradation mechanism decoupling algorithm, and generates a dynamic model of the component-level degradation trajectory and the circuit-level failure propagation path; The adaptive life prediction unit introduces a long short-term memory network and a random forest fusion algorithm based on the dynamic model, learns the correlation features of real-time degradation data and historical failure cases of the circuit through online learning, dynamically corrects the prediction parameters, and outputs the remaining life probability distribution and performance inflection point warning; The intelligent fault decision unit establishes a hierarchical threshold decision mechanism, wherein a dynamic reference library based on circuit function failure definition is arranged at the bottom layer, a fuzzy reasoning engine is used to fuse multiple parameter degradation degrees and failure correlations at the middle layer, and a fault type is analyzed through a fault tree and a Bayesian network coupling model at the top layer, so that progressive diagnosis from weak anomaly to function failure is realized. The early degradation capture unit deploys a high-frequency wide-range sampling circuit and a noise suppression type signal chain, combines time-frequency domain joint feature extraction technology to separate degradation feature signals from background noise, and compares the degradation feature signals with real-time data of the failure mechanism dynamic modeling unit to locate initial degradation elements. The health management collaborative execution unit triggers a maintenance strategy matrix according to the prediction and decision results, performs power flexible derating on a module with a critical life, starts redundancy switching on a branch where a diagnosed degradation element is located, and implements fuse isolation on a sudden fault.

[0007] Compared with the prior art, the present application has the following advantages: The present application has the following advantages: The present application has the following advantages:

[0008] Figure 1 The overall block diagram of the present application is shown in the figure; Figure 2 The load jump transient drop tolerance curve is shown in the figure; Figure 3 The typical direct current system structure diagram is shown in the figure.

[0009] The meanings of the various labels in the figure are as follows: 1, failure mechanism dynamic modeling unit; 2, adaptive life prediction unit; 3, intelligent fault decision unit; 4, early degradation capture unit; 5, health management collaborative execution unit. DETAILED DESCRIPTION

[0010] The technical solutions in the embodiments of the present application will be apparently and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0011] The present application provides a modular DC power supply system with fault self-diagnosis, please refer to Figure 1 The present application provides a modular DC power supply system with fault self-diagnosis, please refer to The failure mechanism dynamic modeling unit 1 associates different environmental stresses and working states to the failure physical model of a specific element by constructing a multi-stress coupling degradation mapping library, separates and superimposes failure effects by using a degradation mechanism decoupling algorithm, and generates a dynamic model of the element-level degradation trajectory and the circuit-level failure propagation path; The adaptive life prediction unit 2 introduces a long short-term memory network and a random forest fusion algorithm based on the dynamic model, dynamically corrects the prediction parameters by learning the associated features of real-time degradation data and historical failure cases of the circuit, and outputs the residual life probability distribution and performance inflection point warning; The intelligent fault decision unit 3 establishes a hierarchical threshold decision mechanism, wherein the bottom layer sets a dynamic reference library based on the functional failure definition of the circuit, the middle layer uses a fuzzy reasoning engine to fuse the multi-parameter degradation degree and failure correlation, and the top layer analyzes the fault type through a fault tree and a Bayesian network coupling model to realize progressive diagnosis from weak anomalies to functional failure; The early degradation capture unit 4 deploys a high-frequency wide-range sampling circuit and a noise suppression type signal chain, separates the degradation feature signal from the background noise by combining time-frequency domain joint feature extraction technology, and locates the initial degradation element by comparing with the real-time data of the failure mechanism dynamic modeling unit (1); The health management collaborative execution unit 5 triggers a preventive maintenance strategy matrix according to the prediction and decision results, performs power flexible derating on the predicted life critical module, starts redundancy switching on the branch where the diagnosed degradation element is located, and implements fuse isolation on the sudden failure.

[0012] When constructing the multi-stress coupling degradation mapping library in the failure mechanism dynamic modeling unit 1, a stress coupling factor matrix and a state influence operator are introduced, wherein the environmental stresses include temperature gradient field, humidity permeability and mechanical vibration spectrum, and the working states cover load jump depth and start-stop cycle cumulative damage amount; By establishing a stress-state joint accelerated degradation test platform, and using nonlinear response surface fitting technology to associate multi-dimensional stress / state variables with the degradation coefficients of the failure physical model, a dynamically updated stress failure mapping matrix is generated.

[0013] The degradation mechanism decoupling algorithm combines intrinsic mode decomposition and fault propagation graph theory, specifically including: By variational modal decomposition, the interference component is stripped, the pure degradation intrinsic mode is extracted, and the failure propagation directed graph dependent on the circuit topology is constructed. The power device is taken as the root node, and the sampling resistor / driving chip is taken as the leaf node. The edge weight is given based on the electrical and thermal coupling strength between nodes. Finally, the degradation energy flow tracking algorithm is used to quantify the transmission path of the failure effect between components, generate the degradation trajectory chain with time sequence label and dynamic fault propagation network, and realize the physical separation of superposition effect.

[0014] The adaptive life prediction unit 2 long short-term memory network and random forest fusion algorithm deploys a double-channel feature extraction architecture, specifically including: The first channel uses the gating mechanism of long short-term memory network to learn the time sequence dependence characteristics of degradation trajectory chain, capture the slowly changing degradation mode, and the second channel analyzes the graph features of fault propagation network through the multi-decision tree structure of random forest, identifies the burst failure correlation, and integrates online incremental learning engine and historical failure case feature clustering engine. The similarity between the real-time collected degradation data and the clustered historical failure case library is matched, and the network weight and forest splitting node are dynamically corrected.

[0015] The dynamic correction prediction parameter adopts a rolling time domain optimization strategy and a confidence interval evaluation mechanism, specifically including: When outputting the residual life probability distribution, a set of random propagation paths of the degradation trajectory chain is generated based on Monte Carlo simulation, the failure probability density function of each edge in the fault propagation network is combined, the confidence interval bandwidth of the life distribution is calculated, and for performance inflection point warning, a gradient mutation detection module and a multi-source evidence fusion module are introduced. When the second derivative of the degradation intrinsic mode exceeds the adaptive threshold and at least two independent sensor channels detect abnormal correlation, a cross-level warning signal is triggered, breaking through the limitation of high false alarm rate of single parameter threshold.

[0016] In the hierarchical threshold decision mechanism of the intelligent fault judgment unit 3, the bottom layer dynamic reference library is built-in circuit function failure rule template, the transient drop tolerance curve under load jump is defined for output voltage tolerance, and the frequency domain harmonic energy threshold surface is defined for ripple coefficient. The middle layer fuzzy reasoning engine constructs the degradation degree and failure correlation tensor model, maps the parameter degradation degree to the membership vector, and outputs the comprehensive degradation index through tensor contraction operation. The top layer fault tree and Bayesian network coupled model adopts dynamic parent node activation strategy. When the comprehensive degradation index exceeds the critical threshold, the corresponding power device failure basic event in the fault tree is activated, and the Bayesian network is driven to infer the potential fault component set.

[0017] The high-frequency wide-range sampling circuit of the early degradation capture unit 4 adopts a modulator and a distributed sampling clock architecture to capture transient current spikes at a sampling rate several times higher than the switching frequency. A noise suppression signal chain design includes a common-mode eddy current cancellation loop and a power supply ground isolation loop. The common-mode noise cancellation circuit board ground is injected with an anti-phase common-mode noise to eliminate the interference. A time-frequency domain joint feature extraction technology implements adaptive wavelet packet decomposition. The feature frequency band is separated through the optimal wavelet packet basis screening. The time-frequency distribution energy of the selected frequency band is focused to extract the film crack acoustic emission characteristic frequency of the metallized film capacitor from the background noise.

[0018] The extracted film crack acoustic emission characteristic frequency is matched with the degradation trajectory chain in the failure mechanism dynamic modeling unit 1 to lock the position of the acoustic emission source corresponding to the electrolytic capacitor. Based on the edge weight direction of the fault propagation directed graph, a correlation backtracking algorithm is started to backtrack the associated node parameters. When the gate voltage oscillation amplitude of the driving chip and the capacitor degradation degree satisfy the preset coupling equation, it is determined that the capacitor is the initial degradation element.

[0019] The power flexible derating of the health management cooperative execution unit 5 is performed on the predicted life critical module. A multi-level power threshold is constructed based on the life and load mapping model. When the lower limit of the life confidence interval is lower than the preset safety baseline, the derating strategy matrix is activated. A dynamic derating slope control technology is used to adjust the output power drop gradient according to the life decay rate. A ripple compensation algorithm is introduced to suppress the output voltage fluctuation caused by derating.

[0020] According to the fault element set output by the intelligent fault judgment unit 3, the branch number of the fault element in the circuit topology is located. When the branch health degree is lower than the health degree of the redundant module, a multi-channel multiplexing switching circuit based on solid-state relays is triggered. The branch health degree is defined as the remaining life of the weakest element in the branch. For sudden faults, a fault type directional fusing strategy is used. For short-circuit faults, a microsecond-level magnetic latching fuse is started. For open-circuit faults, a capacitive energy storage auxiliary shutdown is enabled, and a breakdown arc suppression circuit is combined to eliminate the isolation transient overvoltage.

[0021] Further explanation is needed. The specific implementation of the hierarchical threshold decision mechanism of the intelligent fault judgment unit 3 is as follows: after the remaining life probability distribution output by the adaptive life prediction unit 2 and the performance inflection point warning, the intelligent fault judgment unit 3 needs to locate the fault from weak anomalies to functional failure through the hierarchical threshold decision mechanism, avoiding false positives or false negatives caused by single-level judgment. The hierarchical threshold decision mechanism of the intelligent fault decision unit 3 gradually progresses from the bottom layer to the top layer. The bottom layer dynamic benchmark library is built-in circuit function failure rule templates. The circuit function failure rule templates are a set of fault judgment rules preset based on the core function of the DC power supply, including parameter normal range, abnormal threshold, and failure determination standard. For example, an output voltage deviation of ±5% from the rated value is abnormal, and a deviation of ±10% is failure. The template can dynamically load corresponding rules according to the power supply model. The transient drop tolerance curve under load step change is defined for the output voltage tolerance. The specific definition process is as follows: As shown in Figure 2 , first set the load step change amplitude interval, such as 0-100% rated load, divided into 5 levels at 20% step size. For each step change amplitude, test the maximum allowed transient drop value of the power supply through experiments, such as 0.5V drop allowed when the step change is 20%, and 1.2V drop allowed when the step change is 100%. Take the load step change amplitude as the horizontal axis and the allowed drop value as the vertical axis. Fit a smooth curve, such as an exponential curve. The larger the step change, the more the allowed drop value increases in proportion. The area below the curve is the tolerable area. If it exceeds, it is determined to be voltage abnormality. Define the frequency domain harmonic energy threshold surface for the ripple coefficient. The definition process is as follows: Divide the ripple frequency into 0-1kHz, 1-10kHz, and 10-100kHz three frequency bands. For each frequency band, test the maximum allowed harmonic energy under different loads. Take load rate-frequency-harmonic energy as three-dimensional coordinate axes. Fit the allowed energy values of each test point into a continuous surface. The points on the surface are the upper limit of harmonic energy under the corresponding working condition. If it exceeds, it is determined to be ripple abnormality. The middle layer fuzzy reasoning engine constructs a degradation degree and failure correlation tensor model. The construction process is as follows: First, the model dimensions are determined using a three-dimensional tensor with parameter type (output voltage, ripple, temperature), degradation degree (0-1, 0 for normal, 1 for failure), and failure correlation (correlation with other parameters, 0-1). Historical degradation data is collected, such as a degradation degree of 0.2 when the voltage drops from 12V to 11.5V, and a correlation with ripple of 0.6. The values ​​of each coordinate point in the tensor are filled in, and the parameter degradation degree is mapped to a membership vector. The parameter degradation degree is calculated using real-time collected parameter values ​​and thresholds from a dynamic benchmark library. For example, if the current voltage is 11.7V and the rated voltage is 12V... The degradation degree is calculated as (12 - 11.7) / (12 × 10%) = 0.25. The membership vector converts the degradation degree into the membership degree of a fuzzy set ("normal", "slightly abnormal", "severely abnormal", "failed"). For example, a degradation degree of 0.25 corresponds to a membership degree of 0.6 for "normal" and 0.4 for "slightly abnormal". The comprehensive degradation index is output through tensor shrinking operation. Tensor shrinking operation sums the failure correlation dimensions of the tensor and then weights it with the membership vector. The weights are set according to the importance of the parameters: voltage 0.4, ripple 0.3, and temperature 0.3. For example, voltage membership... After weighting the degree vector, the ripple is 0.18, the temperature is 0.15, and the comprehensive degradation index is 0.24 + 0.18 + 0.15 = 0.57. A higher index indicates a higher failure risk. The top-level fault tree and Bayesian network coupled model employs a dynamic parent node activation strategy. This model combines the basic event-intermediate event-top event structure of the fault tree with the probabilistic reasoning of the Bayesian network. The dynamic parent node activation strategy means that the parent node (basic event) of the corresponding parameter in the fault tree is activated only when the comprehensive degradation index of a certain parameter exceeds a threshold. Power The basic events of device failure are the events at the lowest level of the fault tree. When the comprehensive degradation index exceeds the critical threshold, the system automatically activates the basic events associated with the index in the fault tree, driving the Bayesian network to infer the set of potential faulty components. The Bayesian network inverse reasoning uses the known occurrence probabilities of basic events to infer the posterior probability of each component failure. For example, after activating the capacitor leakage basic event, the network calculates that capacitor C1 has a failure probability of 85% and C2 has a failure probability of 10%, and outputs C1 with the highest probability as the potential faulty component, thus achieving accurate positioning from the abnormal index to the specific component.

[0022] The specific implementation of the early degradation capture unit is as follows: After the failure mechanism dynamic modeling unit 1 generates the component degradation trajectory, the early degradation capture unit 4 needs to separate weak degradation signals from the background noise to avoid the early degradation signals being masked and thus causing missed detection. The high-frequency wide-range sampling circuit of the early degradation capture unit 4 adopts a modulator and distributed sampling clock architecture. The modulator is a Σ-Δ modulator (16-bit resolution), which can convert analog signals into high-resolution digital signals and reduce quantization noise. The distributed sampling clock architecture deploys independent clock modules at key power supply nodes (such as the output terminals of power devices and across the sampling resistors), synchronized by the master clock (synchronization error <10ns), avoiding the delay deviation of a centralized clock. The sampling rate is set to 5 times the power supply switching frequency; for example, if the switching frequency is 100kHz, the sampling rate is 500kHz. A sampling rate higher than the switching frequency can completely capture transient current spikes. The 2μs current spike during MOSFET switching prevents high-frequency transient signals from being masked by sampling aliasing. A noise-suppressing signal chain design incorporates a common-mode eddy current cancellation loop and a power-to-ground isolation loop. The common-mode eddy current cancellation loop uses a reverse coil wound at the signal chain input. When common-mode noise, such as electromagnetic interference from surrounding equipment, occurs at frequencies of 50Hz / 100Hz, the reverse coil generates opposite eddy currents, canceling the magnetic field effect of the original eddy currents. The power-to-ground isolation loop divides the circuit board into independent power-to-ground and signal-to-ground areas, connected by a ferrite bead, to suppress high-frequency noise and prevent ground bounce interference caused by power-to-ground current fluctuations. Ground bounce interference refers to instantaneous changes in ground potential. To offset the signal reference, ground bounce interference is further canceled by injecting inverse common-mode noise. At the amplifier input of the signal chain, noise with the same amplitude but opposite phase to the ground bounce interference is injected; for example, if the ground bounce interference is +0.1V, -0.1V noise is injected, reducing the noise at the final input amplifier to below 0.01V, ensuring signal purity. Adaptive wavelet packet decomposition is implemented using time-frequency domain joint feature extraction technology. Adaptive wavelet packet decomposition automatically selects the decomposition level based on the signal's time-frequency characteristics. For example, if the degraded signal is concentrated in the 1-10kHz range, it is decomposed into 4 levels to avoid feature loss caused by fixed decomposition. Feature frequency bands are separated through optimal wavelet packet basis selection. The optimal basis selection adopts the minimum entropy criterion. The smaller the entropy value, the more concentrated the features. From the multiple frequency bands after decomposition, 3-5 frequency bands with the smallest entropy values ​​are selected. These frequency bands contain the acoustic emission signals of the metallized film capacitor membrane crack. Then, time-frequency distribution energy focusing is performed on the selected frequency bands. The Wigner-Ville distribution (time-frequency analysis method) is used to focus the signal energy in the frequency band to a specific point in the time-frequency plane. The characteristic frequency of membrane crack acoustic emission is extracted from the background noise (energy dispersion). When the metallized film capacitor membrane cracks, it will generate an acoustic emission signal of about 2.5kHz. The energy is concentrated at this frequency point, realizing the separation of weak degradation features.

[0023] The specific implementation method for locating the initial degradation component is as follows: After extracting the characteristic frequency of acoustic emission from the film crack of the metallized film capacitor, it is necessary to locate the specific initial degradation component by matching it with the degradation data of the failure mechanism dynamic modeling unit 1, so as to avoid misjudging other components. The specific implementation method is as follows: First, the extracted acoustic emission characteristic frequency of the membrane crack (e.g., 2.5kHz) is matched with the degradation trajectory chain in the failure mechanism dynamic modeling unit 1. The degradation trajectory chain is the component degradation time series generated by unit 1. For example, the acoustic emission frequency of the membrane crack of capacitor C1 gradually increases from 2.0kHz to 2.5kHz, corresponding to a degradation degree increasing from 0.3 to 0.6. The similarity between the real-time extracted frequency and the frequency of each capacitor degradation trajectory chain is calculated. For example, the similarity with C1 is 0.92, and the similarity with C2 is 0.3. The capacitor (C1) with a similarity > 0.8 is the electrolytic capacitor corresponding to the acoustic emission source, thus initially locking the position of C1. The coordinates on the circuit board are X1, Y1. Based on the edge weight direction of the fault propagation directed graph, the correlation backtracking algorithm is started to backtrack and detect the parameters of the associated nodes. The fault propagation directed graph is the circuit topology constructed by the failure mechanism dynamic modeling unit 1. The root node is the MOS transistor, and the leaf nodes are capacitors and resistors. The edge weight reflects the electrothermal coupling strength between nodes. For example, the edge weight between C1 and the driver chip U1 is 0.7. This indicates strong coupling. The correlation backtracking algorithm starts from the locked C1 and backtracks along the direction with an edge weight > 0.5 (pointing to U1), detecting the associated parameters of U1, such as the gate voltage oscillation amplitude. The normal oscillation amplitude of the gate voltage of the driver chip U1 is 0.2V. When C1 degrades, the capacitance value decreases, causing the gate voltage oscillation amplitude of U1 to increase, such as to 0.5V. When the gate voltage oscillation amplitude of the driver chip and the degree of capacitor degradation satisfy the preset coupling equation, the capacitor is determined to be the initial degraded element. The preset coupling equation is a quantitative relationship established through experiments, such as gate voltage oscillation amplitude = 0.2V + 0.5 × degree of capacitor degradation. For example, the current degree of degradation of C1 is 0.6. Substituting into the equation, we get the theoretical oscillation amplitude = 0.2 + 0.5 × 0.6 = 0.5V, which is consistent with the actual detected 0.5V. In addition, the parameters of other associated nodes (such as sampling resistor R1) are normal (voltage drop of 0.1V, within the normal range). Then, C1 is finally determined to be the initial degraded element, avoiding misjudgment due to a single parameter abnormality.

[0024] The specific implementation method of power flexible derating of the health management collaborative execution unit is as follows: When the adaptive life prediction unit 2 outputs a critical life warning of the module, such as when the lower limit of the remaining life confidence interval is <30 days, the health management collaborative execution unit 5 needs to perform power flexible derating to avoid accelerated degradation of the module due to overload, while ensuring output stability. The specific implementation method is as follows: The health management collaborative execution unit 5 first constructs a multi-level power threshold based on the lifetime and load mapping model. This model is an output power-remaining lifetime relationship model established through accelerated degradation testing. For example, when the output power is 100% of the rated value, the remaining lifetime is 30 days; when it is 80%, the remaining lifetime is 60 days. The multi-level power thresholds are divided according to the remaining lifetime; for example, 30-60 days corresponds to 90% rated power, 15-30 days corresponds to 80%, and <15 days corresponds to 70%. The threshold values ​​are stored in the derating strategy matrix, where rows correspond to lifetime intervals and columns correspond to power values. Based on the remaining lifetime probability distribution output by the adaptive lifetime prediction unit 2, such as a 90% confidence interval of 25-40 days, when the lower limit of the lifetime confidence interval (25 days) is lower than the preset safety baseline (30 days), the derating strategy matrix is ​​activated, retrieving the power threshold (80% rated power) for the corresponding lifetime interval (15-30 days). Dynamic derating slope control technology is then used, based on the lifetime decay rate. The output power derating gradient is adjusted, and the dynamic derating slope is the power reduction rate set according to the lifetime decay rate. For example, if the lifetime decays by 1 day per day, the power decreases by 1% of the rated value per day. This avoids load power instability caused by sudden derating. For example, if the current power is 100% and the target power is 80%, with a lifetime decay rate of 1 day / day, it will take 20 days to reach the target power, decreasing by 1% per day (100%→99%→...→80%). At the same time, a ripple compensation algorithm is introduced to suppress output voltage fluctuations caused by derating. The ripple compensation algorithm monitors the output ripple in real time during power derating. If the ripple increases from 0.1V to 0.2V after derating, the feedback loop parameters of the power supply are adjusted, such as increasing the charging current of the compensation capacitor, to suppress the ripple back to within 0.1V. The algorithm uses proportional-integral (PI) control and dynamically adjusts the compensation amount according to the ripple deviation (actual ripple - target ripple) to ensure that the output voltage fluctuation during derating is <±0.05V, meeting the power supply accuracy requirements of the load.

[0025] The specific implementation of redundancy switching and sudden fault isolation for faulty branches is as follows: After the intelligent fault judgment unit 3 outputs the set of faulty components, the health management collaborative execution unit 5 needs to prevent the fault from spreading and ensure the continuous operation of the power system through redundancy switching and fuse isolation. First, based on the set of faulty components output by the intelligent fault judgment unit 3 (such as capacitor C1 and resistor R2), locate their branch numbers in the circuit topology (e.g., C1 and R2 are located in branch 3). Calculate the health of this branch, which is defined as the remaining lifespan of the weakest component within the branch (e.g., C1 has a remaining lifespan of 10 days, R2 has a remaining lifespan of 20 days, branch health = 10 days). Compare the branch health with the health of the redundant module (e.g., redundant branch health is 60 days). When the branch health (10 days) is lower than the redundant module health... At 60 days, a multiplexing switching circuit based on solid-state relays is triggered. Solid-state relays are electronic switches without mechanical contacts (response time < 1ms). The multiplexing switching circuit includes switching logic for the main branch and redundant branches. During switching, the solid-state relay of the redundant branch is closed first (connected to the redundancy module). After the output stabilizes (about 5ms), the relay of the faulty branch is opened to achieve seamless switching and avoid load power failure. When implementing fuse isolation for sudden faults (such as short circuits and open circuits), a fault type-oriented fuse strategy is adopted. For short-circuit faults (such as a sudden current surge to 10 times the rated value due to MOSFET breakdown), a microsecond-level magnetic latching fuse is activated. This magnetic latching fuse is a fast-acting fuse triggered by electromagnetic force (fusing time < 1μs). When a short-circuit current (e.g., > 5A) is detected, the electromagnetic coil generates a magnetic field, pushing the fuse to open and cutting off the faulty branch. For open-circuit faults (such as current interruption caused by a broken sampling resistor), capacitive energy storage is activated for auxiliary shutdown. An energy storage capacitor (100μF) is connected in parallel in the branch. When an open circuit is detected, the energy storage... The capacitor discharges, providing a brief current to maintain the load power supply (approximately 10ms), while simultaneously triggering the solid-state relay to switch to the redundant branch, preventing the load from crashing due to current interruption. Combined with the breakdown arc suppression circuit, it eliminates isolated transient overvoltages. The breakdown arc suppression circuit consists of a varistor connected in parallel across the fuse (conducting when the voltage exceeds twice the rated value). When the fuse breaks and generates an arc (transient overvoltage of 3 times the rated value), the varistor conducts, absorbing the overvoltage energy and suppressing the voltage to within 1.5 times the rated value, protecting other components from impact.

[0026] In this invention, the failure mechanism dynamic modeling unit 1 constructs a multi-stress coupling degradation mapping library, uses a degradation mechanism decoupling algorithm to separate superimposed failure effects, and generates a dynamic degradation model. The early degradation capture unit 2 deploys a high-frequency wide-area sampling circuit and a noise suppression signal chain, and combines time-frequency domain joint feature extraction to separate degradation feature signals from background noise. The adaptive lifetime prediction unit 3 uses a long short-term memory network and random forest fusion algorithm to dynamically correct parameters and output the remaining lifetime probability distribution. The intelligent fault judgment unit 4 realizes progressive diagnosis through a hierarchical threshold decision mechanism. The health management collaborative execution unit 5 triggers pre-maintenance strategies, executes power flexible derating, redundancy switching, or directional fuse isolation to improve system reliability.

[0027] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A modular DC power supply system with self-diagnostic faults, characterized in that, include: The failure mechanism dynamic modeling unit (1) constructs a multi-stress coupling degradation mapping library, associates different environmental stresses and working states with the failure physical model of a specific component, and uses a degradation mechanism decoupling algorithm to separate superimposed failure effects, generating a dynamic model of component-level degradation trajectory and circuit-level failure propagation path. The adaptive lifetime prediction unit (2) introduces a long short-term memory network and random forest fusion algorithm based on a dynamic model. It dynamically corrects the prediction parameters by learning the correlation characteristics between real-time degradation data of the circuit and historical failure cases online, and outputs the remaining lifetime probability distribution and performance inflection point warning. The intelligent fault judgment unit (3) establishes a hierarchical threshold decision mechanism. The bottom layer sets up a dynamic benchmark library based on the definition of circuit function failure. The middle layer uses a fuzzy inference engine to integrate the multi-parameter degradation degree and failure correlation. The top layer analyzes the fault type through the fault tree and Bayesian network coupling model to realize the progressive diagnosis from weak anomaly to functional failure. The early degradation capture unit (4) deploys a high-frequency wide-area sampling circuit and a noise-suppressed signal chain. Combined with time-frequency domain joint feature extraction technology, it separates degradation feature signals from background noise and locates the initial degradation element by comparing the real-time data with the failure mechanism dynamic modeling unit (1). The health management collaborative execution unit (5) triggers the pre-maintenance strategy matrix based on the prediction and judgment results, performs power derating on the critical modules of the predicted lifespan, initiates redundancy switching on the branch where the diagnosed degraded components are located, and implements fuse isolation for sudden faults.

2. The modular DC power supply system for self-diagnosis of faults according to claim 1, characterized in that: When constructing the multi-stress coupling degradation mapping library in the failure mechanism dynamic modeling unit (1), stress coupling factor matrix and state influence operator are introduced, where stress includes temperature gradient field, humidity permeability and mechanical vibration spectrum, and state covers load jump depth and start-stop cycle cumulative damage. An accelerated degradation test platform combining stress and state was established. The degradation coefficient of the accelerated degradation test platform was correlated with the degradation coefficient of the failure physics model by nonlinear response surface fitting, generating a stress failure mapping matrix, which was used to construct a multi-stress coupled degradation mapping library.

3. The modular DC power supply system for self-diagnosis of faults according to claim 1, characterized in that: The degradation mechanism decoupling algorithm employs a combination of intrinsic mode decomposition and fault propagation graph theory, specifically including: By stripping away interference components through variational mode decomposition, extracting pure degenerate intrinsic modes, and constructing a circuit topology-dependent directed graph of failure propagation, with power devices as root nodes and sampling resistors or driver chips as leaf nodes, edge weights are assigned based on the electrothermal coupling strength between leaf nodes. Finally, the propagation path of failure effects between components is quantified through a degenerate energy flow tracing algorithm, generating a time-marked degenerate trajectory chain and a dynamic fault propagation network to achieve physical separation of superposition effects.

4. The modular DC power supply system for self-diagnosis of faults according to claim 1, characterized in that: The adaptive lifetime prediction unit (2) integrates a long short-term memory network with a random forest algorithm, deploying a dual-channel feature extraction architecture, specifically including: The first channel utilizes the gating mechanism of Long Short-Term Memory (LSTM) networks to learn the temporal dependency characteristics of degradation trajectory chains and obtain slowly changing degradation patterns. The second channel analyzes the graph features of the fault propagation network through the multi-decision tree structure of random forests, identifies sudden failure associations, and integrates an online incremental learning engine and a historical failure case feature clustering engine. It performs similarity matching between the real-time collected degradation data and the clustered historical failure case library, dynamically correcting network weights and forest split nodes.

5. A modular DC power supply system for self-diagnosis of faults according to claim 4, characterized in that: The dynamic correction of prediction parameters employs a rolling time-domain optimization strategy and a confidence interval evaluation mechanism, specifically including: When outputting the remaining lifetime probability distribution, a set of random propagation paths for the degradation trajectory chain is generated based on Monte Carlo simulation. Combined with the failure probability density function of each edge in the fault propagation network, the confidence interval bandwidth of the lifetime distribution is calculated. For performance inflection point warning, a gradient mutation detection module and a multi-source evidence fusion module are introduced. When the second derivative of the degradation intrinsic mode exceeds the adaptive threshold and at least two independent sensor channels detect abnormal correlation, a cross-level warning signal is triggered, overcoming the limitation of high false alarm rate of single parameter threshold.

6. A modular DC power supply system for self-diagnosis of faults according to claim 1, characterized in that: In the hierarchical threshold decision mechanism of the intelligent fault judgment unit (3), the bottom dynamic benchmark library has built-in circuit function failure rule templates, defines the transient drop tolerance curve under load jump for the output voltage tolerance, defines the frequency domain harmonic energy threshold surface for the ripple coefficient, the middle layer fuzzy inference engine constructs the degradation degree and failure correlation tensor model, maps the parameter degradation degree to the membership vector, and outputs the comprehensive degradation index through tensor shrinking operation. The top layer fault tree and Bayesian network coupling model adopts the dynamic parent node activation strategy. When the comprehensive degradation index exceeds the critical threshold, the corresponding power device failure basic event in the fault tree is activated, driving the Bayesian network to reverse infer the potential fault component set.

7. A modular DC power supply system for self-diagnosis of faults according to claim 3, characterized in that: The high-frequency wide-range sampling circuit of the early degradation capture unit (4) adopts a modulator and distributed sampling clock architecture to capture transient current spikes at a sampling rate several times higher than the switching frequency. The noise-suppressed signal chain design uses a common-mode eddy current cancellation loop and a power supply-ground isolation loop to cancel the circuit board ground bounce interference by injecting anti-phase common-mode noise. The time-frequency domain joint feature extraction technology implements adaptive wavelet packet decomposition, and the characteristic frequency band is separated by wavelet packet optimal basis screening. Then, the selected frequency band is subjected to time-frequency distributed energy focusing to extract the membrane crack acoustic emission characteristic frequency of the metallized film capacitor from the background noise.

8. A modular DC power supply system for self-diagnosis of faults according to claim 7, characterized in that: The extracted acoustic emission characteristic frequency of the membrane crack is matched with the degradation trajectory chain in the failure mechanism dynamic modeling unit (1) to lock the position of the electrolytic capacitor corresponding to the acoustic emission source. Based on the edge weight direction of the fault propagation directed graph, the correlation backtracking algorithm is started to backtrack and detect the parameters of the associated nodes. When the gate voltage oscillation amplitude of the driving chip and the degradation degree of the capacitor satisfy the preset coupling equation, the capacitor is determined to be the initial degradation element.

9. A modular DC power supply system for self-diagnosis of faults according to claim 1, characterized in that: In the health management collaborative execution unit (5), the power flexible derating is performed on the critical life prediction module. Based on the life and load mapping model, a multi-level power threshold is constructed. According to the remaining life probability distribution output by the adaptive life prediction unit (2), when the lower limit of the life confidence interval is lower than the preset safety baseline, the derating strategy matrix is ​​activated. The dynamic derating slope control technology is adopted to adjust the output power derating gradient according to the life decay rate. At the same time, the ripple compensation algorithm is introduced to suppress the output voltage fluctuation caused by derating.

10. A modular DC power supply system for self-diagnosis of faults according to claim 9, characterized in that: According to the set of faulty components output by the intelligent fault judgment unit (3), the branch number in the circuit topology is located. When the health of the branch is lower than the health of the redundant module, the multiplexing switching circuit based on solid-state relay is triggered. The health of the branch is defined as the remaining life of the weakest component in the branch. When implementing fuse isolation for sudden faults, the fault type directional fuse strategy is adopted. For short-circuit faults, the microsecond-level magnetic latching fuse is activated. For open-circuit faults, the capacitive energy storage auxiliary shutdown is enabled. The transient overvoltage of the isolation is eliminated in combination with the breakdown arc suppression circuit.

Images