Device and method for evaluating health state of direct-current distribution board

Abstract

The invention discloses a device and a method for evaluating the health state of a direct-current distribution board. The device comprises a sensor layer, a data acquisition instrument, a high-performance computing workstation and an application platform. The method is based on multi-source sensing data and an artificial intelligence algorithm, and comprises the following steps: S1, collecting multi-source data in real time; s2, data edge preprocessing and feature extraction; s3, depth feature learning and preliminary diagnosis on the edge / work station; s4, integrating evaluation and health state output on the work station / platform; and S5, performing result display and early warning on the application platform. The evaluation method comprises a multi-source data acquisition module, an edge feature extraction module, a deep learning diagnosis module and an integrated evaluation module, can perform real-time monitoring and accurate early warning on the health state of the DC distribution board, realizes fusion analysis of multi-physical quantity information of electricity, heat, force and sound, improves the accuracy and reliability of system evaluation, and improves the safety of system evaluation. The operation and maintenance cost is reduced, and safe and stable operation of equipment is guaranteed.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring and fault diagnosis technology, and in particular to a method and device for assessing the health status of DC switchboards. Background Technology

[0002] Currently, the health status assessment of DC distribution boards faces multiple challenges. Traditional methods rely on manual inspections or single-parameter monitoring, making it difficult to detect latent faults such as contact oxidation and loose connections in real time. Existing systems mostly employ offline analysis, requiring data to be uploaded to remote processing, resulting in significant response delays and failing to meet the millisecond-level early warning requirements for rapid faults such as DC arcing. Some intelligent models (such as shallow neural networks) have weak feature extraction capabilities for nonlinear degradation processes such as capacitor aging and insulation deterioration, exhibiting poor generalization and susceptibility to noise interference. Furthermore, the lack of effective fusion mechanisms for multi-source heterogeneous data such as current, voltage, temperature, vibration, and infrared thermography restricts the accuracy of the assessment.

[0003] Among the existing related patent technologies, such as the health status assessment method and device for pole-mounted switchgear based on hybrid intelligent algorithms disclosed in patent (CN110009208A), there is an over-reliance on offline information and historical fault data. The dynamic acquisition and real-time processing capabilities of real-time operating data are insufficient, and it is impossible to achieve "real-time assessment" of health status. It can only complete post-event or phased assessment. The model is only a simple weighted combination of traditional machine learning algorithms. It does not introduce the ability of deep learning to extract high-dimensional and complex features, and has weak adaptability to new and complex faults. It is difficult to cope with multi-parameter coupled fault scenarios of DC distribution boards.

[0004] Therefore, the next generation of technology urgently needs to break through bottlenecks and develop intelligent diagnostic systems that integrate multi-source information, support real-time edge computing, and have strong generalization capabilities, so as to realize the transformation from "passive monitoring" to "active prediction", improve the comprehensiveness, real-time performance and reliability of diagnosis, and ensure the safe and stable operation of DC systems. Summary of the Invention

[0005] This invention provides a method and corresponding hardware device for assessing the health status of DC distribution boards based on multi-source sensor data and artificial intelligence algorithms. It achieves high-precision acquisition of parameters such as current, voltage, temperature, partial discharge, and vibration through multiple sensors. Combining embedded preprocessing and edge intelligent analysis, feature extraction and preliminary assessment are completed at the acquisition end, reducing latency. By integrating ANN, deep learning, and ensemble learning methods, a robust assessment model is constructed, improving the ability to identify complex faults and early degradation, achieving efficient and accurate health monitoring and early warning. This method and device can automatically and in real-time monitor multiple source parameters, including electrical and environmental parameters, during the operation of DC distribution boards, accurately assess their health status, identify hidden faults (such as high-resistance grounding and harmonic interference), and maintain assessment accuracy even when the power grid topology changes dynamically, providing immediate basis for operation and maintenance decisions.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: A DC power distribution board health status assessment device includes a sensor layer, a data acquisition unit, a high-performance computing workstation, and an application platform. The sensor layer comprises various sensors deployed on the DC power distribution board for collecting raw operational data. The data acquisition unit hardware includes a multi-channel ADC, an embedded processor, a communication interface, and supports HTTP / MQTT protocols. The data acquisition unit software includes a data preprocessing algorithm and a lightweight ANN feature extraction model, responsible for data standardization, preliminary feature extraction and compression, and uploading feature data / preliminary results via HTTP / MQTT protocols. The high-performance computing workstation is equipped with a GPU accelerator card for DL ​​model inference and training. The high-performance computing workstation software runs a deep learning (DL) diagnostic model and an ensemble learning evaluation model, receives feature data from the data acquisition unit, performs deep feature learning and multi-model ensemble, and completes the final health status assessment calculation. The application platform receives the workstation's assessment results, provides data storage and visualization, including a health status dashboard, historical trends, alarm logs, early warning management, and report generation functions. It can be accessed via a web interface or API.

[0007] The sensor layer includes various sensors such as current / voltage, temperature, vibration, partial discharge, and infrared sensors. Among them, the DC current sensor is installed on the output busbar of the battery cluster, with a sampling rate of 5kHz and a measurement range of ±2000A. The fiber optic temperature sensor is embedded in the circuit breaker contact, with an accuracy of ±0.1℃ and a response time of <100ms. The wideband vibration sensor is installed on the PCS interface with a bandwidth of 0.5-10kHz.

[0008] The embedded processor uses an ARM Cortex-A series processor; the ANN feature extraction model is a TensorFlow Lite model; and the communication interface is an Ethernet RS485 interface.

[0009] A method for assessing the health status of a DC distribution board, employing the aforementioned assessment device and based on multi-source sensor data and artificial intelligence algorithms, includes the following steps: S1, Real-time acquisition of multi-source data: Real-time acquisition of operational data is achieved through a sensor network deployed at key nodes of the DC distribution board, including input / output buses, switch contacts, capacitor banks, and connectors, serving as the raw signal; S2, Data edge preprocessing and feature extraction: In the data acquisition instrument, the acquired raw signal undergoes preprocessing such as filtering, noise reduction, and normalization. An artificial neural network (ANN)-based feature extractor is applied to reduce the dimensionality of the preprocessed high-dimensional time-series data, extracting key spatiotemporal feature vectors, including harmonic features, temperature gradients, vibration spectrum features, and discharge pulse statistics; S3, Deep feature learning and preliminary diagnosis on the edge / workstation: The extracted feature vectors are input into a deep learning DL model (1D-CNN for waveform features, LSTM) deployed on a high-performance computing workstation or enhanced data acquisition instrument. / GRU is used for time-series dependencies), the DL model learns the complex relationships between features, and outputs preliminary diagnostic results / probability vectors for specific components or fault modes (such as contact overheating, capacitance decay, insulation abnormality); S4, Integrated assessment and health status output on workstation / platform: The feature vectors extracted by ANN and the preliminary diagnostic results of the DL model are input into the ensemble learning model (XGBoost, LightGBM, Random Forest or Stacking ensemble). The ensemble model integrates information from multiple models to perform the final health status assessment and outputs the overall health score: 0-100%, the health status of key components (normal / attention / abnormal / severe) and the probability of potential fault types; S5, Result display and early warning on the application platform: The assessment results are pushed to the application platform in real time for visualization (dashboard, trend chart). When the health score is lower than the threshold or the probability of a specific fault mode exceeds the set value, a graded early warning (such as alarm, trip suggestion) is triggered.

[0010] Furthermore, in step S1, a multi-type sensor network synchronously collects raw signals of multiple physical quantities such as current, voltage, temperature, vibration, and partial discharge to build a comprehensive data sensing foundation.

[0011] Furthermore, in step S2, at the edge, various signals are first filtered, denoised, and normalized, and then key features are extracted dimensionally: For a current signal, calculate the DC ripple rate.

[0012] in, The DC component is filtered using a 100ms moving average, used to reflect the degree of capacitor aging; it also detects sudden current changes.

[0013] When the empirical coefficient , When the current is at the rated current, it is judged as a risk of poor contact and marked with a sudden change indicator. ); For a temperature signal, calculate the rate of temperature rise.

[0014] It is used to predict hotspot development and construct temperature distribution entropy based on multi-point temperature data to assess the uniformity and anomalies of the heat dissipation system.

[0015] in, For the number of sensors, This is the total temperature value; In vibration signal processing, by analyzing vibration acceleration signals... Perform Fourier transform Extract the main resonant frequency

[0016] Used for diagnosing structural faults such as mechanical loosening, all features are extracted and compressed at the edge, significantly reducing the transmission load while retaining fault-sensitive information, providing high-quality input for subsequent in-depth diagnosis and integrated evaluation.

[0017] Furthermore, in step S3, the multi-source features extracted in S2 are used to construct a 5-dimensional input vector:

[0018] These vectors represent the capacitor's aging degree, temperature rise rate, heat dissipation uniformity, dominant mechanical vibration frequency, and current surge frequency, respectively. This vector is input to a lightweight feedforward neural network with 5 input nodes and 16 hidden nodes using the ReLU activation function.

[0019] And three output nodes using the Sigmoid activation function,

[0020] The network outputs three compressed features with clear physical meanings: F1 is the contact degradation index: 0 indicates excellent condition, and 1 indicates severe degradation, used to evaluate the contact performance of contacts or connectors; F2 is the capacitor aging factor: 0 indicates brand new, and 1 indicates near failure, reflecting the health status of filter capacitors; F3 is the thermal runaway risk probability: ranging from 0 to 1, characterizing the likelihood of overheating failure in the device. This process runs efficiently on edge devices, compressing high-dimensional features into three highly interpretable indicators, significantly reducing the burden of subsequent transmission and computation, while retaining key fault information, providing reliable input for subsequent in-depth diagnostics.

[0021] Furthermore, in step S4, deep learning diagnosis is performed on a high-performance workstation, and a dual-path fusion model architecture is adopted to realize multi-dimensional fault identification. Among them, the time-series diagnosis branch is based on an LSTM network and receives 120-second sliding window data composed of contact degradation index F1, capacitor aging factor F2 and average temperature T to capture the time-series evolution characteristics of the equipment status.

[0022] Furthermore, in step S4, a dynamic weighted ensemble evaluation is performed to synthesize the diagnostic results and output a device health score. First, the feature fusion input is a three-dimensional probability vector from an LSTM branch. Two-dimensional probability vectors of the 1D-CNN branch And additional feature F3, represented as , Next, the XGBoost decision model is used for ensemble learning, and the objective function is optimized using the following formula:

[0023] in Used to control model complexity and ensure that the model neither overfits nor underfits. In addition, to improve the model's adaptability and accuracy, a dynamic weighting mechanism was introduced, assigning weights based on the accuracy of each sub-model over the past 24 hours:

[0024] Finally, the health score is calculated based on the prediction results of the ensemble model, using the following formula:

[0025] Formula (14) maps the probability values ​​of poor contact, overheating, mechanical loosening, and additional feature F3 to a percentage-based health score, thus intuitively reflecting the current health status of the equipment. The entire process achieves effective integration and accurate evaluation of multi-source data through dynamic adjustment of weights and optimization of the objective function.

[0026] Furthermore, the method also includes linkage control to ensure stable operation and fault response, taking corresponding control actions to adjust the operating status of the microgrid based on different ranges of equipment health scores.

[0027] The beneficial effects of this invention are: The real-time health status assessment method for DC power distribution boards of the present invention includes a multi-source data acquisition module, an edge feature extraction module, a deep learning diagnosis module, and an integrated assessment module. It can monitor and provide accurate early warning of the health status of DC power distribution boards in real time, realize the fusion analysis of multiple physical quantities such as electricity, heat, force, and sound, improve the accuracy and reliability of system assessment, reduce operation and maintenance costs, and ensure the safe and stable operation of equipment. Attached Figure Description

[0028] Figure 1 This is a flowchart of the DC power distribution board health status assessment method and corresponding hardware device system of the present invention. Detailed Implementation

[0029] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0030] Example: In accordance with the invention described in the "A Method for Evaluating the Health Status of a DC Distribution Board Based on Multi-Source Sensor Data and Artificial Intelligence Algorithms and a Corresponding Hardware Device System", a method for evaluating the health status of a DC distribution board based on multi-source sensor data and artificial intelligence algorithms and a corresponding hardware device system are designed and developed. Figure 1 This describes the specific processing flow of the method, which mainly includes the following steps: (S1) The hardware system consists of a sensing layer, an edge computing layer, and a central decision-making layer.

[0031] Specifically, the DC current sensor in the sensing layer is installed on the output busbar of the battery cluster, with a sampling rate of 5kHz and a measurement range of ±2000A; the fiber optic temperature sensor is embedded in the circuit breaker contact, with an accuracy of ±0.1℃ and a response time of <100ms; the wideband vibration sensor is installed on the PCS interface, with a bandwidth of 0.5-10kHz, etc.; the intelligent data acquisition device in the edge computing layer adopts an ARM Cortex-A53 processor and an integrated real-time signal processing module, and supports HTTP / MQTT dual protocol transmission; the AI ​​workstation in the central decision-making layer includes an NVIDIA GPU accelerator and runs a three-level AI evaluation model.

[0032] (S2) Multi-source data acquisition and edge preprocessing.

[0033] The system utilizes a multi-type sensor network deployed at key locations (such as busbar connection points, switch contacts, capacitor banks, and terminal blocks) to simultaneously collect raw signals of various physical quantities, including current, voltage, temperature, vibration, and partial discharge, thus constructing a comprehensive data sensing foundation. At the edge, various signals are first filtered, denoised, and normalized, and then key features are extracted dimensionally. For the current signal, the DC ripple rate is calculated.

[0034] in, The DC component is filtered using a 100ms moving average, used to reflect the degree of capacitor aging; it also detects sudden current changes.

[0035] When the empirical coefficient , When the current is at the rated current, it is judged as a risk of poor contact and marked with a sudden change indicator. ).

[0036] For a temperature signal, calculate the rate of temperature rise.

[0037] It is used to predict hotspot development and construct temperature distribution entropy based on multi-point temperature data to assess the uniformity and anomalies of the heat dissipation system.

[0038]

[0039] in, For the number of sensors, This represents the total temperature value.

[0040] In vibration signal processing, by analyzing vibration acceleration signals... Perform Fourier transform Extract the main resonant frequency

[0041] This technology is used to diagnose structural faults such as mechanical loosening. All features are extracted and compressed at the edge, significantly reducing transmission load while retaining fault-sensitive information, providing high-quality input for subsequent in-depth diagnosis and integrated evaluation.

[0042] (S3) Edge-side ANN feature compression and key state index extraction.

[0043] The multi-source features extracted from (S1) are used to construct a 5-dimensional input vector:

[0044] These parameters represent the capacitor's aging degree, temperature rise rate, heat dissipation uniformity, dominant mechanical vibration frequency, and current surge frequency, respectively. This vector is input to a lightweight feedforward neural network, which consists of 5 input nodes and 16 hidden nodes using the ReLU activation function.

[0045] And three output nodes using the Sigmoid activation function,

[0046] The network outputs three compressed features with clear physical meanings: F1 is the contact degradation index (0 indicates excellent condition, 1 indicates severe degradation), used to evaluate the contact performance of contacts or connectors; F2 is the capacitor aging factor (0 indicates brand new, 1 indicates near failure), reflecting the health status of the filter capacitor; and F3 is the thermal runaway risk probability (range 0~1), characterizing the likelihood of the device experiencing an overheating failure. This process runs efficiently on edge devices, compressing high-dimensional features into three highly interpretable indicators, significantly reducing the burden of subsequent transmission and computation, while retaining key fault information, providing reliable input for subsequent in-depth diagnostics.

[0047] (S4) Perform deep learning diagnostics on the workstation In (S3), the system performs deep learning diagnostics on a high-performance workstation, employing a dual-path fusion model architecture to achieve multi-dimensional fault identification. The time-series diagnostic branch, based on an LSTM network, receives a 120-second sliding window of data consisting of the contact degradation index F1, capacitor aging factor F2, and average temperature T, capturing the temporal evolution characteristics of the equipment state. This branch models long-term dependencies through a gating mechanism; its cell state updates are controlled by an input gate, a forget gate, and an output gate.

[0048] Final cell state:

[0049] Hidden state:

[0050] Output three-dimensional probability vectors for poor contact, overheating, and normal conditions. The spectral diagnostic branch employs a 1D-CNN network, taking a 1000-point normalized vibrational spectrum envelope within the 500Hz frequency band as input. It extracts local frequency domain features through multi-layer convolution operations, with the convolution output being...

[0051] The characteristics of mechanical loosening and resonance risks are abstracted layer by layer, and the corresponding two-dimensional probability vector is finally output. The two branches process data in parallel, focusing on electrical thermal temporal behavior and mechanical vibration spectrum characteristics respectively, enabling refined diagnosis of various types of faults and providing complementary evidence for subsequent integrated decision-making.

[0052] (S5) Dynamic weighted integrated evaluation The system performs a dynamic weighted ensemble evaluation to synthesize diagnostic results and output a device health score. First, the feature fusion input consists of a three-dimensional probability vector from an LSTM branch. Two-dimensional probability vectors of the 1D-CNN branch And additional feature F3, represented as .

[0053] Next, the XGBoost decision model is used for ensemble learning. The objective function is optimized using the following formula:

[0054] in Used to control model complexity and ensure that the model is neither overfitted nor underfitted.

[0055] In addition, to improve the model's adaptability and accuracy, a dynamic weighting mechanism was introduced, assigning weights based on the accuracy of each sub-model over the past 24 hours:

[0056] Finally, the health score is calculated based on the prediction results of the ensemble model, using the following formula:

[0057] This formula maps the probability values ​​of poor contact, overheating, mechanical loosening, and the additional feature F3 to a percentage-based health score, thus intuitively reflecting the current health status of the equipment. The entire process achieves effective integration and accurate assessment of multi-source data through dynamic adjustment of weights and optimization of the objective function.

[0058] (S6) Linkage Control Interlocking control is used to ensure stable operation and fault response. Based on different ranges of equipment health scores, corresponding control actions are taken to adjust the operating status of the microgrid.

Claims

1. A direct current distribution board health state evaluation device, characterized by, The system comprises a sensor layer, a data acquisition instrument, a high-performance computing workstation, and an application platform. The sensor layer includes various sensors deployed on the DC distribution panel to collect raw operation data. The data acquisition instrument hardware has multi-channel ADC, embedded processor, communication interface, and supports HTTP / MQTT protocol. The data acquisition instrument software includes operation data preprocessing algorithm and lightweight ANN feature extraction model, which is responsible for data standardization, preliminary feature extraction and compression, and uploads feature data / preliminary results through HTTP / MQTT protocol. The high-performance computing workstation is equipped with GPU acceleration card hardware facilities for DL model inference and training. The high-performance computing workstation software runs deep learning (DL) diagnostic model and ensemble learning evaluation model, receives feature data from the data acquisition instrument, performs deep feature learning and multi-model integration, and completes the final health state evaluation calculation. The application platform receives workstation evaluation results, provides data storage and visual display, including health state dashboard, historical trend, alarm log, early warning management, report generation, and other functions. It can be accessed through a Web interface or API.

2. The DC power distribution panel health assessment device of claim 1, wherein, The various sensors in the sensor layer include current / voltage, temperature, vibration, partial discharge, and infrared sensors. The DC current sensor is installed on the battery cluster output busbar with a sampling rate of 5 kHz and a measurement range of ±2000A. The optical fiber temperature sensor is embedded in the circuit breaker contact with an accuracy of ±0.1℃ and a response time of <100ms. The wideband vibration sensor is installed on the PCS interface with a bandwidth of 0.5-10kHz.

3. The DC power distribution panel health assessment device of claim 1, wherein, The embedded processor uses ARM Cortex-A series processor. The ANN feature extraction model is TensorFlow Lite model. The communication interface is Ethernet RS485 interface.

4. A method for health state assessment of a DC distribution board, using the assessment device of any one of claims 1-3, based on multi-source sensing data and artificial intelligence algorithms, comprising the following steps: S1, Real-time collection of multi-source data: Through the deployment of sensor networks on key nodes of the DC distribution panel, including input / output busbars, switch contacts, capacitor banks, and connection bars, real-time collection of operation data as raw signals; S2, Data edge preprocessing and feature extraction: In the data acquisition instrument, the collected raw signals are preprocessed by filtering, noise reduction, and normalization. The ANN-based feature extractor is applied to reduce the dimensionality of the preprocessed high-dimensional time series data and extract key spatiotemporal feature vectors, including harmonic features, temperature gradient, vibration spectrum features, and discharge pulse statistics; S3, Deep feature learning and preliminary diagnosis on the edge / workstation: The extracted feature vectors are input into the deep learning (DL) model deployed in the high-performance computing workstation or the enhanced data acquisition instrument. The DL model learns the complex relationships between features and outputs preliminary diagnosis results / probability vectors for specific components or fault modes; S4, Ensemble evaluation and health state output on the workstation / platform: The feature vectors extracted by ANN and the preliminary diagnosis results of the DL model are input into the ensemble learning model. The ensemble model integrates multi-model information and performs final health state evaluation, outputting overall health score: 0-100%, key component health state, and potential fault type probability. S5, result display and early warning on application platform: the evaluation results are pushed to the application platform in real time, and visualized display is performed, when the health score is lower than the threshold value or the probability of detecting a specific failure mode exceeds the set value, a hierarchical early warning is triggered.

5. The DC power distribution panel health assessment method of claim 4, wherein, In step S1, a multi-type sensor network is used to synchronously collect current, voltage, temperature, vibration, partial discharge and other multi-physical quantity original signals, and a comprehensive data sensing basis is constructed.

6. The DC power distribution panel health assessment method of claim 4, wherein, In step S2, at the edge side, first, various signals are filtered, denoised and normalized, and then key features are extracted in different dimensions: For current signals, the DC ripple rate is calculated, wherein, is the 100 ms moving average filtered DC component, used to reflect the degree of capacitor aging; simultaneously detect current mutations, When the experience coefficient , is the rated current, it is determined that there is a risk of poor contact and a mutation flag is marked ( ), For temperature signals, the temperature rise rate is calculated, which is used to predict the development of hot spots, and a temperature distribution entropy is constructed based on multi-point temperature data to evaluate the uniformity and abnormality of the heat dissipation system, wherein, is the number of sensors, is the total temperature value, In the vibration signal processing, the main resonance frequency is extracted by performing Fourier transform on the vibration acceleration signal ​​ which is used to diagnose structural failures such as mechanical looseness, and all features are extracted and compressed on the edge side, significantly reducing transmission load while retaining fault-sensitive information, providing high-quality input for subsequent deep diagnosis and integrated evaluation.

7. The DC power distribution panel health assessment method of claim 4, wherein, In step S3, the multi-source features extracted in S2 are constructed into a 5-dimensional input vector: representing the capacitance aging degree, temperature rise rate, heat dissipation uniformity, mechanical vibration main frequency and current mutation frequency, this vector is input into a lightweight feedforward neural network, the network structure includes 5 input nodes, 16 hidden nodes using ReLU activation function, and 3 output nodes using Sigmoid activation function, the network outputs three compressed features with clear physical meaning: F1 is the contact degradation index: 0 represents good state, 1 represents serious degradation, used to evaluate the contact performance of contacts or connectors; F2 is the capacitance aging factor: 0 represents new, 1 represents close to failure, reflecting the health status of the filter capacitor; F3 is the thermal runaway risk probability: range 0~1, representing the possibility of overheating failure of the device. This process runs efficiently on the edge device, compressing high-dimensional features into 3 high-interpretability indicators, significantly reducing subsequent transmission and calculation burden while retaining key fault information, providing reliable input for subsequent deep diagnosis.

8. The DC power distribution panel health assessment method of claim 4, wherein, In step S4, deep learning diagnosis is performed on a high-performance workstation, and a dual-path fusion model architecture is used to realize multi-dimensional fault recognition, wherein the time series diagnosis branch is based on the LSTM network and receives 120-second sliding window data composed of the contact degradation index F1, the capacitance aging factor F2 and the average temperature T to capture the time evolution characteristics of the device state.

9. The DC power distribution panel health assessment method of claim 4, wherein, In step S4, a dynamic weighted ensemble evaluation is performed to integrate the diagnostic results and output the device health score. First, the feature fusion input is composed of the three-dimensional probability vector from the LSTM branch , the two-dimensional probability vector from the 1D-CNN branch , and the additional feature F3, denoted as , Next, an XGBoost decision model is used for integrated learning, and the objective function optimization is realized by the following formula: wherein for controlling the complexity of the model, ensuring that the model is neither overfit nor underfit, In addition, in order to improve the adaptability and accuracy of the model, a dynamic weight mechanism is introduced, and the weights are allocated based on the accuracy of each sub-model in the past 24 hours: Finally, the health score output is calculated according to the prediction result of the integrated model, and the specific formula is as follows: This formula (14) maps the probability values of contact failure, overheating, mechanical looseness and additional feature F3 to a health score in percentage, thereby directly reflecting the current health status of the device. Through dynamic adjustment of weights and optimization of objective functions, the process realizes effective integration and accurate evaluation of multi-source data.

10. The DC power distribution panel health assessment method of any of claims 4-9, wherein, The method further comprises linkage control for ensuring stable operation and fault response, and according to different intervals of the equipment health score, corresponding control actions are taken to adjust the operation state of the microgrid.

Images