Modular monitoring and fault pre-diagnosis system for direct current power supply system

Abstract

The application belongs to the technical field of power system monitoring, and specifically discloses a DC power supply system modular monitoring and fault pre-diagnosis system. The system comprises an edge intelligent agent unit arranged at each power supply module and a collaborative diagnosis server at the system level. The edge unit performs high-frequency data acquisition and local health assessment, the collaborative server fuses multi-module data, dynamically constructs a fault propagation graph based on a coupling relationship model, and uses a long short-term memory network for trend prediction and fault pre-diagnosis. The application realizes double-layer health assessment and early warning from the module to the system, and improves the fault positioning accuracy and system reliability.

Description

Technical Field

[0001] This invention belongs to the field of power system monitoring technology, specifically relating to a modular monitoring and fault pre-diagnosis system for DC power supply systems. Background Technology

[0002] In the fields of power electronics and industrial automation, DC power supply systems serve as core power supply units, and their reliability and stability directly affect the safe operation of the entire electrical equipment or production line. As system scale and complexity increase, modular DC power supply systems composed of multiple independent power modules connected in parallel or series have become the mainstream architecture to achieve power expansion, redundancy backup, and flexible configuration.

[0003] Status monitoring and fault diagnosis of DC power supply systems are key technologies for ensuring their long-term reliable operation. This technology aims to assess the system's health by collecting electrical parameters and operating status of each module in real time, and to provide early warnings and location of faults before or at the initial stage, thereby avoiding systemic downtime caused by single-point failures.

[0004] Existing technologies typically configure an independent monitoring unit for each power module, collecting basic data such as voltage, current, and temperature and uploading it to a central controller for centralized display and threshold alarms. However, this centralized monitoring architecture has the following problems: the monitoring data of each module is isolated from each other, lacking in-depth analysis and modeling of the electrical and thermal coupling relationships between modules, resulting in an unclear system-level fault evolution mechanism. When a module experiences a latent fault or performance degradation, its impact will propagate throughout the system through current sharing loops, bus voltage, and other pathways. Existing methods struggle to quickly trace and locate the root cause module from the complex alarm information.

[0005] Traditional threshold alarms are reactive and cannot intelligently learn and predict the gradual changes in module parameters, thus failing to achieve true pre-fault diagnosis. This results in low troubleshooting efficiency for maintenance personnel when faced with alarms, and an inability to proactively intervene before faults cause substantial impact, leading to persistently high system operational risks. Summary of the Invention

[0006] The purpose of this invention is to provide a modular monitoring and fault pre-diagnosis system for DC power supply systems, in order to solve the technical contradictions in the existing centralized monitoring architecture, such as isolated module status data, inability to analyze the coupling effects between modules, and lack of fault prediction capabilities based on trend learning, which leads to difficulties in locating the root cause of faults and delayed early warning.

[0007] This invention provides a modular monitoring and fault pre-diagnosis system for DC power supply systems. The system is built on a modular DC power supply system containing at least two power modules that operate in parallel or series. It includes an edge intelligent agent unit deployed locally in each power module, a collaborative diagnostic server deployed at the system level, and a data communication network connecting the units. The edge intelligent agent unit is physically integrated inside the corresponding power module and is used to perform local data acquisition, feature extraction and preliminary health assessment. The collaborative diagnostic server is used to receive and integrate data uploaded by all edge intelligent agent units, and perform system-level coupling relationship analysis and fault propagation simulation. The edge intelligent agent unit includes a sensing acquisition subunit, an embedded feature calculation subunit, and a first health assessment subunit; The collaborative diagnostic server includes a data fusion and storage subunit, a system coupling relationship modeling subunit, a fault propagation map construction subunit, and a second health assessment subunit.

[0008] Preferably, the sensing and acquisition subunit synchronously acquires the output voltage, output current, junction temperature of key power devices, and surface temperature of the heat sink of its power module at a sampling frequency greater than 1 kHz. The embedded feature calculation subunit processes the collected raw time-series data in real time and extracts multi-dimensional feature vectors including the effective value of voltage ripple, current dynamic regulation rate, temperature rise slope, and amplitude of harmonic components in a specified frequency band. The first health assessment subunit has a built-in benchmark health model trained based on the historical full life cycle data of the power module of this model. The benchmark health model defines the normal fluctuation range of each feature vector under standard operating conditions. The first health assessment subunit inputs the multidimensional feature vector calculated in real time into the benchmark health model and outputs a local health index with a value range of 0 to 100.

[0009] Preferably, the data fusion and storage subunit receives and time-aligns the local health index, multi-dimensional feature vectors, and original sampled data fragments of all power modules to form a system-level spatiotemporal correlation database; The system coupling relationship modeling subunit is based on circuit topology parameters and thermal design parameters. It constructs and updates a mathematical model offline to describe the electrical and thermal coupling relationship between modules online. The model is specifically characterized as the transfer function matrix of the influence of the output current disturbance of any power module on the output voltage of other modules, and the influence coefficient matrix of the heat dissipation of any power module on the temperature rise of its neighboring modules. The fault propagation graph construction subunit dynamically generates a fault propagation graph in the form of a directed graph based on the mathematical model output by the system coupling relationship modeling subunit and the real-time monitored abnormal module information. The nodes of the directed graph represent each power module, and the directed edges represent the possible propagation paths of faults or performance degradation. The weight of the edges is determined by the strength coefficient of the coupling relationship model and the local health index of the current abnormal module. The second health assessment subunit performs a system-level health status assessment by combining the local health index and fault propagation map information. Its assessment logic is as follows: modules whose current local health index is less than a preset first threshold are marked as suspected abnormal sources. Traverse the fault propagation graph and calculate the potential impact of the suspected anomaly source on other modules via the coupling path; calculate the comprehensive health index for each module by weighting and summing the local health index with all received potential impacts.

[0010] Preferably, the collaborative diagnostic server further includes a trend prediction and fault pre-diagnosis engine; The trend prediction and fault pre-diagnosis engine extrapolates and predicts the key degradation characteristics of each power module based on long-term historical data accumulated in the data fusion and storage sub-unit. The trend prediction and fault pre-diagnosis engine adopts a time-series prediction algorithm based on long short-term memory network, which uses the feature vector sequence of each module in the past continuous time as input to predict the evolution trajectory of the feature vector of the future module. The trend prediction and fault pre-diagnosis engine compares the predicted trajectory with the baseline trajectory at the beginning of the power module's lifespan and calculates the trajectory deviation. When the deviation of the predicted trajectory of any module exceeds the preset second threshold for three consecutive prediction cycles, or when the predicted trajectory shows that its local health index will drop below the preset fault threshold within the next 8 hours, the trend prediction and fault pre-diagnosis engine generates a pre-diagnosis alarm for the power module and estimates its remaining effective running time.

[0011] Preferably, the data interaction between the edge intelligent agent unit and the collaborative diagnostic server adopts a layered communication strategy; The edge intelligent agent unit uploads its calculated local health index and compressed feature vector summary to the collaborative diagnostic server. When the local health index is less than a preset first threshold, the edge intelligent agent unit automatically shortens the communication cycle to 100 milliseconds and uploads a detailed diagnostic data packet containing high-sampling-rate raw data fragments. After receiving the detailed diagnostic data packet, the collaborative diagnostic server instructs the fault propagation map construction subunit to initiate a deep analysis process for the power module.

[0012] Preferably, the online update mechanism for the system coupling relationship modeling subunit is as follows: The system automatically initiates a full-system small-signal disturbance test each time a power module is put into operation, removed, or replaced. During the test, the collaborative diagnostic server sends instructions to the designated edge intelligent agent unit through the data communication network, so that it controls the power module to generate a step disturbance of output current with controllable amplitude and short duration. At the same time, all edge intelligent agent units synchronously monitor and record the voltage and current response of this module; The system coupling modeling subunit collects all response data and recalculates the transfer function matrix of electrical coupling and the influence coefficient matrix of thermal coupling through a system identification algorithm.

[0013] Preferably, when generating the fault propagation map construction subunit, the historical fault records and maintenance logs of the module are also introduced as prior knowledge for weight correction. For modules that have historically caused system-level problems due to specific types of failures, when they are anomaly source nodes in the graph, the weight of their output edges is multiplied by an enhancement factor greater than 1. For modules that have recently undergone preventative maintenance or component replacement, when they are affected nodes, the weights of their input edges are multiplied by a decay factor of less than 1.

[0014] Preferably, the trend prediction and fault pre-diagnosis engine further includes an early warning confidence assessment module; The early warning confidence assessment module calculates a confidence score for each generated pre-diagnostic alarm; The confidence score is calculated based on the completeness of the historical data sequence on which the prediction is based, the recent prediction accuracy of the long short-term memory network model on the validation set, and the degree of matching between the deviation of the current predicted trajectory and the characteristics of the previous stage of similar historical failures. The system will only elevate the pre-diagnostic alert to the highest priority and trigger the proactive intervention process when the confidence score is greater than 75%.

[0015] Preferably, the specific process by which the second health assessment subunit calculates the comprehensive health index for each module is as follows: The comprehensive health index is derived by weighting and summing the local health index with all potential impacts received.

[0016] Preferably, the multidimensional feature vector extracted by the embedded feature calculation subunit is constructed by calculating the effective value of the ripple of the output voltage signal, calculating the dynamic adjustment rate of the output current signal, calculating the rise slope of the temperature signal, and extracting the amplitude of the harmonic components in the specified frequency band by performing a fast Fourier transform on the voltage and current signals.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by deploying edge intelligent agent units in each power module, achieves edge-based status awareness and primary diagnosis, significantly improving the real-time performance and granularity of data acquisition. The collaborative diagnostic server focuses on the fusion of multi-module data and system-level coupled analysis, constructing a two-layer health assessment system from "individual modules" to "the entire system." This system architecture fundamentally changes the isolated data situation in traditional monitoring, enabling the system to accurately quantify the chain reaction of individual module anomalies on the global system through electrical and thermal coupling, thereby achieving rapid and accurate location of the root cause of the fault and shortening the fault diagnosis time.

[0018] 2. This invention innovatively introduces a dynamic fault propagation graph construction mechanism based on a physical model. This mechanism transforms the abstract coupling relationships between modules into a visualized, weighted directed graph model, clearly revealing the potential propagation paths of faults within the system and quantitatively assessing the vulnerability of each module by calculating the impact. Combined with weight correction based on historical maintenance data, the diagnostic reasoning process better aligns with the actual operating history and state of the system, improving the accuracy and intelligence of fault diagnosis under complex operating conditions.

[0019] 3. This invention breaks through the passive mode of traditional threshold alarms. By integrating a trend prediction and fault pre-diagnosis engine based on a long short-term memory network, it achieves active learning and proactive prediction of the gradual performance change process of the power module. The system can predict the degradation trajectory of key parameters based on long-term operating data and issue high-confidence pre-diagnosis alarms several hours before a fault occurs. This capability transforms operation and maintenance from "reactive maintenance" to "predictive maintenance," allowing maintenance activities to be scheduled within a plan, avoiding unplanned downtime, greatly improving the reliability and availability of DC power supply system operation, and reducing the total life cycle operation and maintenance cost. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of system-level health assessment based on coupling relationship modeling and graph construction in this invention; Figure 3 This is a logical flowchart of the edge intelligent agent unit performing local data collection and evaluation in this invention; Figure 4 This is a schematic diagram of the hierarchical data interaction between the edge intelligent agent unit and the collaborative diagnostic server in this invention; Figure 5This is a flowchart illustrating the principle and process framework of the trend prediction and fault pre-diagnosis engine in this invention for proactive early warning. Detailed Implementation

[0021] Example 1: Reference Figures 1 to 5 The present invention proposes a modular monitoring and fault pre-diagnosis system for DC power supply systems. This system is built upon a modular DC power supply system consisting of at least two power modules operating in parallel or series. By integrating edge intelligent agent units within each power module and deploying a collaborative diagnostic server at the system level, combined with a high-speed and reliable data communication network, a two-layer health monitoring system is formed, possessing both local real-time sensing capabilities and global coupled analysis capabilities.

[0022] The entire system's workflow begins with each edge intelligent agent unit collecting and initially assessing the operating status of its respective power module at high frequency. Then, the collaborative diagnostic server completes multi-source data fusion, system-level coupling modeling, fault propagation path deduction, and future trend prediction, ultimately achieving accurate location and early warning of potential faults.

[0023] Please refer to the attached document. Figure 3 Each edge intelligent agent unit is physically integrated within its corresponding power module. Its core functions include three sub-stages: high-precision sensing and acquisition, embedded feature calculation, and initial health assessment. The high-precision sensing and acquisition sub-unit synchronously acquires key operating parameters of its power module at a sampling frequency greater than 1 kHz, specifically including output voltage, output current, junction temperature of key power devices, and surface temperature of the heat sink.

[0024] All four types of signals are conditioned by dedicated analog front-end circuitry. Voltage and current signals are processed by isolation amplifiers and anti-aliasing filters before being sent to a high-resolution analog-to-digital converter, while temperature signals are directly read using thermocouples or digital temperature sensors. All raw sampled data is timestamped and buffered in a local circular buffer to ensure timing consistency in subsequent processing.

[0025] The embedded feature computation subunit reads the raw time-series data stream from the circular buffer and performs real-time feature extraction operations.

[0026] For the output voltage signal, its RMS ripple value is calculated, which is the root mean square value of the AC component after removing the DC component within the power frequency cycle. For the output current signal, its dynamic regulation rate is calculated, defined as the ratio of the overshoot of the current response before and after a load step change to the steady-state value. For the temperature signal, the rise slope of the junction temperature of the power device and the surface temperature of the heat sink are calculated, which is the maximum derivative value of the temperature change per unit time. In addition, a fast Fourier transform is performed on the voltage and current signals to extract the weighted sum of the harmonic component amplitudes within a specified frequency band, such as 10 kHz to 100 kHz.

[0027] The five indicators mentioned above together constitute a multidimensional feature vector with a fixed dimension of 5, each dimension corresponding to a specific physical quantity feature. This physical quantity feature vector is updated once per second and serves as the input for the next stage of health assessment.

[0028] The first health assessment subunit contains a benchmark health model trained based on historical full lifecycle data of this power module model. This benchmark health model establishes the statistical distribution characteristics of each feature dimension under standard operating conditions through offline learning of a large number of normal operating samples. Specifically, each feature dimension has dynamic upper and lower limit threshold ranges, which adaptively adjust according to external conditions such as ambient temperature and load rate. When the multi-dimensional feature vectors calculated in real time are input into the benchmark health model, the model first determines whether each dimension exceeds the normal fluctuation range under the current operating conditions. If any dimension exceeds the limit, a local anomaly marker is triggered.

[0029] Subsequently, the model calculates the local health index at that moment using the weighted Euclidean distance method, based on the degree of violation, duration, and correlation between features. The local health index is strictly limited to a value between 0 and 100, where 100 represents a completely healthy state and 0 represents a severe fault state. This local health index, as the core output of the edge side, is periodically uploaded to the collaborative diagnostic server.

[0030] Please refer to the attached document. Figure 4 The data interaction between the edge intelligent agent unit and the collaborative diagnostic server adopts a hierarchical communication strategy. Under normal system operation, the edge intelligent agent unit uploads its calculated local health index and compressed feature vector summary to the collaborative diagnostic server every 1 second via the data communication network. The feature vector summary is generated by performing principal component analysis on the original five-dimensional feature vector to reduce dimensionality, retaining only the first two principal components, thus minimizing communication bandwidth consumption.

[0031] When the local health index falls below a preset first threshold of 80, the edge intelligent agent unit automatically shortens the communication cycle to 100 milliseconds and initiates the upload mechanism for detailed diagnostic data packets. This data packet contains high-sampling-rate raw data segments, specifically complete raw voltage, current, and temperature sampling sequences within 500 milliseconds before and after the trigger moment, with the sampling rate maintained at 1 kHz. Upon receiving such a high-priority data packet, the collaborative diagnostic server immediately instructs the fault propagation map construction subunit to initiate a deep analysis process for this power module model.

[0032] The collaborative diagnostic server, serving as the central processing node of the system, comprises a data fusion and storage subunit, a system coupling relationship modeling subunit, a fault propagation graph construction subunit, a second health assessment subunit, and a trend prediction and fault pre-diagnosis engine. The data fusion and storage subunit is responsible for receiving data streams from all edge intelligent agent units and performing strict time alignment operations.

[0033] Because each edge unit may experience microsecond-level clock drift, the sub-units employ a synchronization correction algorithm based on a network time protocol to uniformly map all data onto the system's global time axis. The aligned data is categorized and stored in a system-level spatiotemporal relational database according to module identifiers. The database adopts a columnar storage structure, supporting efficient querying and backtracking of feature vectors, health indices, and original data fragments of any module within any time period.

[0034] The system coupling relationship modeling subunit is the core foundation for realizing system-level diagnostic capabilities. Based on the circuit topology parameters and thermal design parameters of the modular DC power supply system, this subunit offline constructs a mathematical model describing the electrical and thermal coupling relationships between modules.

[0035] The electrical coupling model is specifically represented as a square matrix of electrical coupling transfer functions corresponding to the total number of power supply modules currently in operation. The elements at the corresponding positions in the matrix represent the impact of the output current disturbance of one module on the output voltage of another module.

[0036] The thermal coupling model is represented as a square matrix of thermal coupling influence coefficients corresponding to the total number of power supply modules currently in operation. The elements at the corresponding positions in the matrix represent the contribution coefficient of a module's unit heat dissipation to the temperature rise of its neighboring modules. This contribution coefficient is determined by the physical spacing between modules, the air duct layout, and the material of the heat sink.

[0037] The two models mentioned above are not statically fixed but have online update capabilities. Whenever a module is added, removed, or replaced in the system, the collaborative diagnostic server automatically triggers a system-wide small-signal disturbance test. During the test, the server sends a command to a designated edge intelligent agent unit, causing it to control its power module to generate an output current step disturbance with an amplitude of 5% of the rated current and a duration of 100 milliseconds.

[0038] Meanwhile, all edge intelligent agent units synchronously record the voltage and current response waveforms of their respective modules. After collecting all response data, the system coupling modeling subunit uses a least-squares system identification algorithm to refit the parameters of the electrical coupling transfer function matrix, and updates the thermal coupling influence coefficient matrix by inverting the thermal balance equation, thereby ensuring that the coupling model always remains consistent with the current actual system topology.

[0039] Please refer to the attached document. Figure 2 The fault propagation graph construction subunit dynamically generates a fault propagation graph based on the updated coupling relationship model and real-time monitored abnormal module information. This fault propagation graph is organized in the form of a directed graph, where each node corresponds to a power module, and each directed edge represents a path that a fault or performance degradation may propagate from one module to another.

[0040] Path weight is determined by two parts: In the coupling relationship model, the strength coefficient is taken as the absolute value of the gain of the corresponding electrical coupling transfer function in the dominant frequency band for electrical coupling paths, and as the corresponding thermal coupling influence coefficient for thermal coupling paths. The local health index of the current abnormal source module is calculated as follows: the path weight equals the electrical coupling weight coefficient multiplied by the absolute value of the gain of the electrical coupling transfer function in the dominant frequency band, plus the thermal coupling weight coefficient multiplied by the thermal coupling influence coefficient, and then multiplied by the ratio of (100 minus the local health index of the abnormal source module) to 100. The electrical coupling weight coefficient and the thermal coupling influence coefficient are configurable coefficients used to balance the relative importance of electrical and thermal coupling.

[0041] When the local health index of a module is less than 80, the module is marked as a suspected anomaly source. The graph construction sub-unit starts from this module, traverses all outgoing edges, and calculates its potential impact on downstream modules. The impact is calculated using a path decay accumulation model, specifically: For any module, the degree to which it is affected by a certain abnormal source module is equal to the sum of the path weights of all paths connecting the abnormal source module and the current module, and then multiplied by the attenuation coefficient related to the path length (the longer the path length, the smaller the attenuation coefficient).

[0042] The second health assessment subunit integrates the local health index and fault propagation map information to calculate a comprehensive health index for each module. This index is calculated as follows: the comprehensive health index equals the local assessment weight coefficient multiplied by the module's local health index, plus (1 minus the local assessment weight coefficient) multiplied by (100 minus the sum of the potential impacts of all other modules on this module). The local assessment weight coefficient is a predefined coefficient ranging from 0.6 to 0.8 to ensure that the local status plays a dominant role in the comprehensive assessment.

[0043] The overall health index is also normalized to a range of 0 to 100. The lower the value, the more severe the impact of other abnormal modules, making it the weakest link most likely to experience secondary failures. The collaborative diagnostic server sorts all modules based on the overall health index and lists modules with an overall health index below 70 as high-risk objects, pushing them to the operation and maintenance management interface for focused monitoring.

[0044] The trend prediction and fault pre-diagnosis engine is a key component of this system to achieve proactive maintenance capabilities. Its working principle is shown in the attached figure. Figure 5 As shown, this trend prediction and fault pre-diagnosis engine uses long-term historical data accumulated in the data fusion and storage sub-units to extrapolate and predict the key degradation characteristics of each power module. The input data is a sequence of feature vectors sampled hourly over the past 720 consecutive hours (i.e., 30 days) for each module, totaling 720 time steps.

[0045] The engine employs a time-series prediction algorithm based on a Long Short-Term Memory (LSTM) network. This LSM network consists of two LSM hidden layers, each with 128 neurons, and a fully connected output layer used to predict the feature vector evolution trajectory for each hour within the next 24 hours. During training, the network is optimized by minimizing the mean squared error between the predicted trajectory and the actual observed trajectory, and is periodically fine-tuned online using newly acquired data to adapt to pattern drift caused by module aging.

[0046] After prediction is completed, the engine compares the predicted trajectory with the baseline trajectory established by the module in the early stages of its lifespan (i.e., the first 100 hours after commissioning) and calculates the trajectory deviation. The trajectory deviation is defined as the dynamic time warped distance between the predicted trajectory and the baseline trajectory in the feature space, which is then normalized and mapped to the interval of 0 to 100.

[0047] When the trajectory deviation of any module exceeds the preset second threshold of 85 for three consecutive prediction periods (i.e., 72 hours), or when the predicted trajectory indicates that its local health index will drop below the fault threshold of 60 within the next 8 hours, the engine generates a pre-diagnostic alarm for that module. The alarm content includes a fault type prediction (such as capacitor aging, heat dissipation failure, etc.), the estimated remaining effective operating time, and recommended maintenance measures.

[0048] To enhance the reliability of early warnings, the trend prediction and fault pre-diagnosis engine also includes an early warning confidence assessment module. This module calculates a confidence score for each generated pre-diagnosis alert. The confidence score is calculated based on three core indicators: The confidence level is reduced if the missing rate in the 720-hour data is greater than 5%, based on the completeness of the historical data sequence on which the prediction is based. The average prediction accuracy of the Long Short-Term Memory Network model on the validation set of the most recent 30 days. For every 1 percentage point decrease in accuracy, the confidence level is adjusted accordingly. The matching degree between the currently predicted trajectory deviation pattern and the feature library of similar historical faults is measured by cosine similarity. The three indicators are weighted and fused to generate the final confidence score.

[0049] The system will only elevate the pre-diagnostic alarm to the highest priority and trigger the proactive intervention process when the confidence score is greater than 75%. The proactive intervention process includes: automatically scheduling backup power modules to operate and share the load, pushing maintenance work orders with highlighted marks to the operation and maintenance management system, and notifying on-duty personnel to perform planned maintenance through the human-machine interface, thereby completing the handling before the fault actually occurs and avoiding unplanned downtime.

[0050] Throughout the system's operation, data and control flows between components are highly coordinated. Edge intelligent agent units continuously provide high-granularity state awareness, while collaborative diagnostic servers build system-level cognition on this basis. By deeply integrating physical coupling models, graph theory analysis methods, and deep learning prediction techniques, this system not only solves the problem of isolated module data in traditional monitoring but also achieves a paradigm shift from "passive response" to "proactive prediction."

[0051] The system can accurately identify the chain reaction caused by the deterioration of a single module and provide early warning of potential faults several hours in advance, significantly improving the operational reliability and intelligent operation and maintenance level of the DC power supply system.

[0052] Example 2: Based on the previous examples, this example further refines the thermal coupling model construction method of the system coupling relationship modeling subunit and enhances the depth of utilization of historical maintenance data by the fault propagation map construction subunit.

[0053] Specifically, the initial construction of the thermal coupling influence coefficient matrix no longer relies solely on static thermal design parameters, but instead incorporates on-site measured thermal response data for calibration. After the system's initial commissioning or large-scale upgrade, the collaborative diagnostic server initiates a complete thermal field calibration procedure. This procedure controls each power module to operate stably for 30 minutes at different load rates (e.g., 30%, 60%, 100% of rated power), during which all edge intelligent agent units continuously record the surface temperature changes of the heat sinks of their own module and neighboring modules.

[0054] After collecting these steady-state and transient temperature rise data, the system coupling modeling sub-unit uses the finite difference method to invert and solve the heat conduction equation, thereby obtaining a more accurate inter-module thermal influence coefficient. This process ensures that the thermal coupling model can reflect non-ideal factors such as airflow obstruction and cabinet layout in the actual installation environment.

[0055] Furthermore, when calculating edge weights, the fault propagation graph construction subunit incorporates historical fault records and maintenance logs of modules as prior knowledge for dynamic correction. The system maintenance database stores a complete history of each module, including fault types, occurrence times, root cause analysis reports, and maintenance / replacement records. When a module is marked as a suspected anomaly source, the graph construction subunit first queries its historical records.

[0056] If a module has caused a system voltage collapse due to output capacitor failure within the past 12 months, then when it is considered an anomalous source node, the weights of all outgoing edges will be multiplied by an enhancement factor of 1.3. Conversely, if the module has undergone preventative replacement of critical power devices within the last 30 days, then when it is considered an affected node, the weights of all incoming edges will be multiplied by a reduction factor of 0.7. This correction mechanism ensures that the fault propagation map not only reflects the current physical coupling strength but also incorporates the historical reliability information of individual devices, making diagnostic reasoning more targeted and accurate.

[0057] In terms of trend prediction, this embodiment expands the input features of the Long Short-Term Memory (LSTM) network. In addition to the original five-dimensional feature vector, contextual features such as ambient temperature, load rate fluctuation, and cumulative operating hours are added as auxiliary inputs. Although these features do not directly participate in the local health index calculation, they have a significant impact on the module aging rate.

[0058] The network structure was adjusted accordingly to a multi-input channel. The main channel processes core electrical and thermal features, while the auxiliary channel processes environmental and usage features. The two channels are fused after the long short-term memory hidden layer. This improvement enables the prediction model to distinguish between temporary performance fluctuations caused by changes in external operating conditions and permanent degradation trends caused by internal deterioration, thereby significantly reducing the false alarm rate.

[0059] Through the above enhancement measures, this embodiment exhibits stronger robustness and adaptability in complex industrial environments, and is particularly suitable for large data centers or rail transit power supply systems with frequent changes in module configuration and varied operating conditions.

[0060] 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.

[0061] 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 modular monitoring and fault pre-diagnosis system for DC power supply systems, characterized in that, The system is built on a modular DC power supply system containing at least two power modules that operate in parallel or series, including an edge intelligent agent unit deployed locally in each power module, a collaborative diagnostic server deployed at the system level, and a data communication network connecting the units. The edge intelligent agent unit is physically integrated inside the corresponding power module and is used to perform local data acquisition, feature extraction and preliminary health assessment. The collaborative diagnostic server is used to receive and integrate data uploaded by all edge intelligent agent units, and perform system-level coupling relationship analysis and fault propagation simulation. The edge intelligent agent unit includes a sensing acquisition subunit, an embedded feature calculation subunit, and a first health assessment subunit; The collaborative diagnostic server includes a data fusion and storage subunit, a system coupling relationship modeling subunit, a fault propagation map construction subunit, and a second health assessment subunit.

2. The modular monitoring and fault pre-diagnosis system for DC power supply systems according to claim 1, characterized in that, The sensing and acquisition subunit synchronously acquires the output voltage, output current, junction temperature of key power devices, and surface temperature of the heat sink of its power module at a sampling frequency greater than 1 kHz. The embedded feature calculation subunit processes the collected raw time-series data in real time and extracts multi-dimensional feature vectors including the effective value of voltage ripple, current dynamic regulation rate, temperature rise slope, and amplitude of harmonic components in a specified frequency band. The first health assessment subunit has a built-in benchmark health model trained based on the historical full life cycle data of the power module of this model. The benchmark health model defines the normal fluctuation range of each feature vector under standard operating conditions. The first health assessment subunit inputs the multidimensional feature vector calculated in real time into the benchmark health model and outputs a local health index with a value range of 0 to 100.

3. The modular monitoring and fault pre-diagnosis system for DC power supply systems according to claim 2, characterized in that, The data fusion and storage subunit receives and time-aligns the local health index, multi-dimensional feature vectors, and original sampled data fragments of all power modules to form a system-level spatiotemporal correlation database; The system coupling relationship modeling subunit is based on circuit topology parameters and thermal design parameters. It constructs and updates a mathematical model offline to describe the electrical and thermal coupling relationship between modules online. The mathematical model is specifically characterized as the transfer function matrix of the influence of the output current disturbance of any power module on the output voltage of other modules, and the influence coefficient matrix of the heat dissipation of any power module on the temperature rise of its neighboring modules. The fault propagation graph construction subunit dynamically generates a fault propagation graph in the form of a directed graph based on the mathematical model output by the system coupling relationship modeling subunit and the real-time monitored abnormal module information. The nodes of the directed graph represent each power module, and the directed edges represent the possible propagation paths of faults or performance degradation. The weight of the edges is determined by the strength coefficient of the coupling relationship model and the local health index of the current abnormal module. The second health assessment subunit performs a system-level health status assessment by combining the local health index and fault propagation map information. Its assessment logic is as follows: modules whose current local health index is less than a preset first threshold are marked as suspected abnormal sources. Traverse the fault propagation graph and calculate the potential impact of the suspected anomaly source on other modules via the coupling path; calculate the comprehensive health index for each module by weighting and summing the local health index with all received potential impacts.

4. The modular monitoring and fault pre-diagnosis system for DC power supply systems according to claim 3, characterized in that, The collaborative diagnostic server also includes a trend prediction and fault pre-diagnosis engine; The trend prediction and fault pre-diagnosis engine extrapolates and predicts the key degradation characteristics of each power module based on long-term historical data accumulated in the data fusion and storage sub-unit. The trend prediction and fault pre-diagnosis engine adopts a time-series prediction algorithm based on long short-term memory network, which uses the feature vector sequence of each module in the past continuous time as input to predict the evolution trajectory of the feature vector of the future module. The trend prediction and fault pre-diagnosis engine compares the predicted trajectory with the baseline trajectory at the beginning of the power module's lifespan and calculates the trajectory deviation. When the deviation of the predicted trajectory of any module exceeds the preset second threshold for three consecutive prediction cycles, or when the predicted trajectory shows that its local health index will drop below the preset fault threshold within the next 8 hours, the trend prediction and fault pre-diagnosis engine generates a pre-diagnosis alarm for the power module and estimates its remaining effective running time.

5. The modular monitoring and fault pre-diagnosis system for DC power supply systems according to claim 4, characterized in that, The data interaction between the edge intelligent agent unit and the collaborative diagnostic server adopts a layered communication strategy. The edge intelligent agent unit uploads its calculated local health index and compressed feature vector summary to the collaborative diagnostic server. When the local health index is less than a preset first threshold, the edge intelligent agent unit automatically shortens the communication cycle to 100 milliseconds and uploads a detailed diagnostic data packet containing high-sampling-rate raw data fragments. After receiving the detailed diagnostic data packet, the collaborative diagnostic server instructs the fault propagation map construction subunit to initiate a deep analysis process for the power module.

6. The modular monitoring and fault pre-diagnosis system for DC power supply systems according to claim 5, characterized in that, The online update mechanism for the system coupling relationship modeling subunit is as follows: The system automatically initiates a full-system small-signal disturbance test each time a power module is put into operation, removed, or replaced. During the test, the collaborative diagnostic server sends instructions to the designated edge intelligent agent unit through the data communication network, so that it controls the power module to generate a step disturbance of output current with controllable amplitude and short duration. At the same time, all edge intelligent agent units synchronously monitor and record the voltage and current response of this module; The system coupling modeling subunit collects all response data and recalculates the transfer function matrix of electrical coupling and the influence coefficient matrix of thermal coupling through a system identification algorithm.

7. The modular monitoring and fault pre-diagnosis system for DC power supply systems according to claim 6, characterized in that, When generating the fault propagation graph, the subunit also incorporates the module's historical fault records and maintenance logs as prior knowledge for weight correction. For modules that have historically caused system-level problems due to specific types of failures, when they are anomaly source nodes in the graph, the weight of their output edges is multiplied by an enhancement factor greater than 1. For modules that have recently undergone preventative maintenance or component replacement, when they are affected nodes, the weights of their input edges are multiplied by a decay factor of less than 1.

8. The modular monitoring and fault pre-diagnosis system for DC power supply systems according to claim 7, characterized in that, The trend prediction and fault pre-diagnosis engine also includes an early warning confidence assessment module; The early warning confidence assessment module calculates a confidence score for each generated pre-diagnostic alarm; The confidence score is calculated based on the completeness of the historical data sequence on which the prediction is based, the recent prediction accuracy of the long short-term memory network model on the validation set, and the degree of matching between the deviation of the current predicted trajectory and the characteristics of the previous stage of similar historical failures. The system will only elevate the pre-diagnostic alert to the highest priority and trigger the proactive intervention process when the confidence score is greater than 75%.

9. The modular monitoring and fault pre-diagnosis system for DC power supply systems according to claim 1, characterized in that, The specific process by which the second health assessment subunit calculates the comprehensive health index for each module is as follows: The comprehensive health index is derived by weighting and summing the local health index with all potential impacts received.

10. The modular monitoring and fault pre-diagnosis system for DC power supply systems according to claim 1, characterized in that, The multidimensional feature vector extracted by the embedded feature calculation subunit is constructed by calculating the effective value of ripple on the output voltage signal, calculating the dynamic adjustment rate on the output current signal, calculating the rise slope on the temperature signal, and extracting the amplitude of the harmonic components in the specified frequency band by performing a fast Fourier transform on the voltage and current signals.

Images