10 kilovolt power distribution network fault detection method and system based on Internet of Things

Abstract

The invention provides a 10 kilovolt power distribution network fault detection method and system based on the Internet of Things, and the method comprises the steps: constructing a dynamic digital twinborn body of a power distribution network, verifying an equipment health state index through an annular verification mechanism, and generating a multi-dimensional health portrait; training a fault diagnosis model, and extracting fault features; and constructing a three-level output model, performing end-to-end optimization on the fault feature vector, and generating a diagnosis result including fault type, position and severity evaluation. According to the invention, a multi-dimensional health portrait is generated by using an annular verification mechanism, the equipment health state is comprehensively evaluated, and the fault feature extraction integrity is improved; a fault diagnosis model is trained by means of a longitudinal federated learning architecture, data privacy is protected, and the generalization ability of the model and the adaptability to complex fault scenes are improved; three-level output model optimization is constructed based on a related algorithm and a network, a diagnosis result is accurately generated, and the fault diagnosis accuracy of the 10-kilovolt power distribution network is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid fault detection, in particular to a 10 kilovolt distribution network fault detection method and system based on the Internet of Things. BACKGROUND

[0002] In the field of power system operation and maintenance, 10 kilovolt distribution network as a key link connecting the power transmission network and end users, its stable and reliable operation is crucial to ensure the quality of power supply. With the vigorous development of the Internet of Things technology, its application in distribution network fault detection is increasingly widespread, providing new ideas and means to improve the intelligent level of fault detection. At present, there are many distribution network fault detection technologies based on the Internet of Things. Some methods deploy various sensors on distribution network equipment, real-time collect multi-source data such as electrical parameters, temperature, vibration, and transmit data to the monitoring center for analysis and processing through wired or wireless communication networks. Some technologies use big data analysis to mine massive historical fault data, establish a fault feature library, and identify faults by comparing real-time data with data in the feature library. Some researches use machine learning algorithms such as support vector machine and decision tree to train the collected data and build a fault diagnosis model to realize fault classification and positioning. These existing technologies have improved the efficiency and accuracy of distribution network fault detection to some extent.

[0003] However, the existing 10 kilovolt distribution network fault detection technology based on the Internet of Things still has many shortcomings. First, in terms of data utilization, although multi-source data is collected, there is often a lack of effective data fusion and deep mining methods, making it difficult to fully and accurately reflect the real operating state of the equipment. The correlation and complementarity between different types of data cannot be fully utilized, resulting in insufficient fault feature extraction and affecting the accuracy of fault diagnosis. SUMMARY

[0004] The present application aims to at least solve the technical problems in the prior art that the accuracy of fault diagnosis is low, and particularly innovatively proposes a 10 kilovolt distribution network fault detection method and system based on the Internet of Things.

[0005] In order to achieve the above-mentioned purpose of the present application, the present application provides a 10 kilovolt distribution network fault detection method based on the Internet of Things, the method comprising: S1, constructing a dynamic digital twin of the distribution network; S2, verifying the health state index of the equipment based on the dynamic digital twin using a ring verification mechanism, and generating a multi-dimensional health portrait; S3, training a fault diagnosis model using a vertical federated learning architecture; S4, extracting fault features based on the multi-dimensional health portrait using the fault diagnosis model; S5. Based on a bio-inspired swarm intelligence optimization algorithm and a deep Q-network, a three-level output model is constructed to perform end-to-end optimization of the fault feature vector and generate diagnostic results that include fault type, location and severity assessment.

[0006] On the other hand, the present invention also provides a 10 kV distribution network fault detection system based on the Internet of Things, the system comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement an IoT-based 10 kV distribution network fault detection method when executing executable instructions.

[0007] The beneficial effects of this invention are as follows: By constructing a dynamic digital twin of the distribution network, this invention can accurately simulate the real-time operating status of the distribution network, providing comprehensive and accurate basic data support for fault detection; by utilizing a ring verification mechanism to generate multi-dimensional health profiles, it comprehensively assesses the health status of equipment at multiple scales (equipment, region, system), fully explores data correlations and features, and effectively improves the completeness of fault feature extraction; by using a vertical federated learning architecture to train the fault diagnosis model, it enhances the model's generalization ability by integrating multi-regional data while protecting data privacy, thus enhancing its adaptability to complex fault scenarios; and by constructing a three-level output model based on a bio-inspired swarm intelligence optimization algorithm and a deep Q-network for end-to-end optimization, it can accurately generate diagnostic results that include fault type, location, and severity assessment. This method significantly improves the accuracy of fault diagnosis in 10 kV distribution networks.

[0008] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0009] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of a 10 kV distribution network fault detection method based on the Internet of Things according to the present invention. Detailed Implementation

[0010] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] Example 1 like Figure 1As shown, a fault detection method for a 10 kV distribution network based on the Internet of Things (IoT) includes: S1. Construct a dynamic digital twin of the distribution network; S2. Based on the dynamic digital twin, the ring verification mechanism is used to verify the health status indicators of the equipment and generate a multi-dimensional health profile. S3. Train the fault diagnosis model using a vertical federated learning architecture; S4. Extract fault features based on multi-dimensional health profiles using fault diagnosis models; S5. Based on a bio-inspired swarm intelligence optimization algorithm and a deep Q-network, a three-level output model is constructed to perform end-to-end optimization of the fault feature vector and generate diagnostic results that include fault type, location and severity assessment.

[0012] In this embodiment, the principle of a 10 kV distribution network fault detection method based on the Internet of Things (IoT) is as follows: First, IoT sensors deployed at key nodes of the distribution network (such as transformers, circuit breakers, and line sectionalizing switches) collect equipment operating parameters (such as voltage, current, temperature, vibration, etc.) in real time and transmit the data to edge computing nodes for preprocessing. The edge computing nodes clean, normalize, and perform preliminary feature extraction on the raw data before uploading it to a cloud-based federated learning platform via a secure encrypted channel. The cloud platform, based on a vertical federated learning architecture, trains a global fault diagnosis model by integrating multi-source heterogeneous data while protecting the data privacy of each region. During training, each data-holding node calculates gradients using local data and uploads them in encrypted form. The central server aggregates gradients using a weighted average algorithm to update global model parameters, iteratively optimizing until the model converges. After model training, a multi-dimensional health profile (covering equipment-level electrical parameters, regional load distribution, and system-level topology) generated from a dynamic digital twin is input into the fault diagnosis model. A convolutional neural network extracts hotspot spatial features, a long short-term memory network captures temporal dependencies, and a multi-head self-attention mechanism fuses equipment degradation features, ultimately generating a fused feature vector. The vector is optimized end-to-end through a three-level output model (the feature fusion layer uses gated recurrent units for dimensionality reduction, the decision optimization layer calculates the action Q-value through a deep Q-network, and the diagnostic output layer parses the action sequence), outputting a diagnostic result containing the fault type probability distribution, location coordinate confidence interval, and severity score vector. The severity score is quantified into low, medium, and high levels using a fuzzy logic evaluation system. The reward function dynamically adjusts the weight ratio of classification, localization, and quantization tasks to ensure the model focuses on key indicators at different training stages. This method, through the synergy of digital twins and federated learning, achieves intelligent end-to-end fault diagnosis from data acquisition to result output, significantly improving diagnostic accuracy and robustness in complex power distribution network scenarios.

[0013] As an optional embodiment of the present invention, optionally, constructing a dynamic digital twin of the distribution network in step S1 includes: S101. Establish a three-dimensional geometric model based on laser point cloud data; In step S101, it is necessary to explain in detail that three-dimensional point cloud data of the power distribution network equipment is acquired through laser scanning technology. This data contains precise geometric information of the equipment surface. Using professional 3D modeling software, this point cloud data is processed, including point cloud filtering and denoising, point cloud registration and fusion, to construct a high-precision 3D geometric model of the power distribution network equipment. This model can realistically reflect the physical form and spatial relationships of the equipment. Furthermore, considering the complexity and diversity of power distribution network equipment, targeted modeling methods and parameter settings are required for different types of equipment (such as transformers, circuit breakers, and transmission lines) to ensure the accuracy and reliability of the 3D geometric model. In addition, to achieve dynamic simulation, the physical properties and operating parameters of the equipment, such as material characteristics and electrical parameters, need to be embedded in the 3D geometric model, enabling the model to dynamically respond and change according to different operating conditions.

[0014] S102. Collect multimodal data of the power grid, map the multimodal data into a three-dimensional geometric model, and fuse the multimodal data with Kalman filtering and deep belief network to eliminate noise interference and correct the positioning error of local overheated areas, forming the basic framework of digital twin. In step S102, it is necessary to explain in detail that the multimodal data encompasses electrical quantity data (such as voltage, current, power, etc.), non-electrical quantity data (such as temperature, humidity, vibration, etc.), and equipment status information (such as switch status, protection action signals, etc.). This multi-source heterogeneous data is collected in real time through a sensor network deployed at each node of the distribution network and transmitted to the data processing center via a data interface. At the data processing center, the collected multimodal data is first preprocessed, including data cleaning, missing value imputation, and outlier removal, to improve data quality. Subsequently, the existing Kalman filter algorithm is used to fuse the multimodal data. The Kalman filter can recursively estimate the system state based on the system's dynamic model and observation data, effectively eliminating noise interference in the data. Simultaneously, the fused data is further optimized using an existing deep belief network. Through the nonlinear transformation of multiple layers of neurons, the deep belief network can automatically extract deep features from the data, correct local overheating area positioning errors, and thus form the basic framework of the digital twin. This framework not only includes the static geometric information of the distribution network equipment but also incorporates the dynamic operating data of the equipment.

[0015] S103. Based on the real-time acquisition of modal data, update the basic framework of the digital twin in real time to obtain a dynamic digital twin of the distribution network.

[0016] In step S103, it is necessary to explain in detail that an efficient data update mechanism is required to achieve dynamic updates of the digital twin. Real-time operational data of the distribution network, including electrical and non-electrical quantities and equipment status information, is continuously collected through an IoT sensor network. This data, after preprocessing, is transmitted to the basic framework of the digital twin. Within the framework, existing incremental learning algorithms are used to process the real-time data. This algorithm can integrate new data into the existing model without retraining the entire model, thus achieving dynamic updates to the basic framework of the digital twin. In this way, the digital twin can reflect the actual operating status of the distribution network in real time, including equipment health, load changes, and potential fault risks. Simultaneously, to ensure the accuracy and reliability of the digital twin, it is also necessary to periodically verify and correct it. By comparing and analyzing the digital twin with the actual operating data of the distribution network, deviations and errors in the digital twin can be identified and corrected in a timely manner.

[0017] As an optional embodiment of the present invention, optionally, in step S2, verifying the device health status indicators based on the dynamic digital twin using a ring verification mechanism to generate a multi-dimensional health profile includes: S201. Based on dynamic digital twins, acquire multi-source real-time data streams of the power distribution network and construct a four-dimensional data matrix of electrical, thermal, mechanical, and electromagnetic components; In step S201, it is necessary to explain in detail that the existing dynamic digital twin of the power distribution network is used to acquire multi-source heterogeneous data streams in real time, covering electrical parameters (voltage, current, phase), thermal parameters (equipment surface temperature, ambient temperature), mechanical parameters (vibration frequency, displacement), and electromagnetic parameters (electric field strength, magnetic field strength), through its built-in sensor network and data processing interface. After preliminary cleaning and standardization at the edge computing nodes, these data are mapped onto the three-dimensional geometric model of the digital twin, forming virtual images that correspond one-to-one with the physical equipment. Subsequently, a multimodal data fusion algorithm (such as a spatiotemporal alignment method based on tensor decomposition) is used to align the dispersed electrical, thermal, mechanical, and electromagnetic data according to the time series, constructing a four-dimensional data matrix. The electrical dimension records the real-time power flow and power quality of the equipment; the thermal dimension reflects the heat loss and heat dissipation status of the equipment; the mechanical dimension captures the mechanical stress and vibration characteristics of the equipment; and the electromagnetic dimension characterizes the electromagnetic field distribution and coupling effects of the equipment. This matrix is ​​updated in real time through the simulation engine of the dynamic digital twin, synchronously reflecting the multi-physics coupling state of the physical equipment in the real operating environment. Meanwhile, to ensure the accuracy and integrity of the data matrix, Kalman filtering is used to suppress high-frequency noise, and a deep belief network is used to intelligently complete missing data, ultimately forming a high-fidelity, low-redundancy four-dimensional data base.

[0018] S202. The outer loop Kalman filter is used to perform time-series alignment and power frequency noise filtering on the electrical-thermal-mechanical-electromagnetic four-dimensional data matrix, and the inner loop deep belief network is used to perform feature fusion on the electrical-thermal-mechanical-electromagnetic four-dimensional data matrix to obtain the equipment health status feature vector. In step S202, it is necessary to explain in detail the deep processing of the four-dimensional data matrix using a hierarchical filtering and feature fusion mechanism. The outer-loop Kalman filter recursively estimates the time-series data across the electrical, thermal, mechanical, and electromagnetic dimensions, achieving millisecond-level time alignment using a prediction-correction mechanism while suppressing 50Hz power frequency interference and sensor drift noise. The inner-loop deep belief network employs a five-layer restricted Boltzmann machine stacked structure, automatically extracting cross-modal coupling features through unsupervised pre-training and fine-tuning strategies: the bottom layer network captures transient correlations between the electrical and thermal fields (such as temperature gradients caused by sudden current surges), the middle layer network analyzes the long-term evolution of the mechanical and electromagnetic fields (such as the co-occurrence mode of vibration spectrum and magnetic field distortion), and the top layer network fuses multi-physics spatiotemporal features to form a device health status feature vector. This vector contains 256 feature parameters, covering key indicators such as insulation aging index (based on thermal-electrical coupling features), mechanical fatigue (based on vibration spectrum energy distribution), and electromagnetic interference intensity (based on magnetic field harmonic distortion rate).

[0019] S203. Update the dynamic digital twin using the equipment health status feature vector to generate health status indicators; In step S203, it is necessary to explain in detail that the 256-dimensional equipment health status feature vector extracted in step S202 is input into the simulation engine of the dynamic digital twin, and the physical property parameters in the three-dimensional geometric model are updated through the backpropagation mechanism. Specifically, the conductivity and dielectric constant of the equipment materials are dynamically adjusted according to the insulation aging index, the elastic modulus and damping coefficient of the structural components are corrected according to the mechanical fatigue, and the permeability and eddy current loss parameters of the coil are updated according to the electromagnetic interference intensity. At the same time, in the health assessment module of the digital twin, the insulation aging index is mapped to an insulation integrity score of 0-100%, the mechanical fatigue is converted into a vibration health level of 0-10, and the electromagnetic interference intensity is quantified as the noise threshold exceedance multiple in dBμV units. A comprehensive health status index (HSI) is generated through a weighted fusion algorithm (the weight coefficients are dynamically adjusted according to the equipment type, such as transformers focusing on insulation indicators and circuit breakers focusing on mechanical performance). Its value range is set from 0 (complete failure) to 1 (brand new state), and it is divided into 10 health levels with an interval of 0.1. The updated digital twin not only includes the device's static geometric information but also embeds real-time updated dynamic health status indicators. To ensure the timeliness of indicator updates, a sliding window mechanism (with a window length of 15 minutes) is used to perform rolling analysis on the feature vector. When the HSI value drops by more than 15% within three consecutive windows, an early warning signal is triggered, and a heat map of the device's health status is displayed in real time through the digital twin's visualization interface (red represents high-risk areas, yellow represents warning areas, and green represents normal areas).

[0020] S204. Based on health status indicators, construct a multi-dimensional health profile at three scales: equipment level, regional level, and system level. The multi-dimensional health profile includes electrical health score, hot spot evolution trend, mechanical fatigue level, and insulation performance degradation curve.

[0021] As an optional embodiment of the present invention, optionally, constructing a multi-dimensional health profile based on health status indicators at three scales—device level, region level, and system level—in step S204 includes: S2041. At the equipment level, an electrical health score is calculated based on the electrical parameters in the health status indicators using a weighted scoring algorithm and updated in real time. In step S2041, it is necessary to explain in detail that key electrical parameters such as voltage fluctuation rate, current distortion rate, and power factor deviation are extracted from the health status indicators. The weight of each parameter is determined using the analytic hierarchy process (AHP) (e.g., voltage fluctuation rate weight 0.4, current distortion rate weight 0.35, power factor deviation weight 0.25). The parameter values ​​are mapped to the [0,1] interval using existing fuzzy membership functions; for example, the membership degree is 0.8 when the voltage fluctuation rate exceeds the rated value by 10%. The electrical health score is calculated using a weighted scoring model: Helec = Σ(wi·μi), where wi is the parameter weight and μi is the parameter membership degree. The scoring results are pushed to the equipment monitoring interface of the digital twin in real time and displayed in the form of a dynamic dashboard. A yellow warning is triggered when the score is below a threshold (e.g., 60 points), and a red alarm is triggered when it is below 30 points. Simultaneously, a historical database of electrical parameters is established, and the electrical health score trend for the next 24 hours is predicted using an ARIMA time series model. To address the differences between various equipment types, the scoring algorithm parameters are customized for equipment such as transformers, circuit breakers, and cables. For example, transformers focus on voltage fluctuation rate and harmonic content, while circuit breakers focus on peak opening and closing current and operating time deviation.

[0022] S2042. At the regional level, integrate health data from multiple devices and use thermal imaging analysis models to generate hot spot evolution trends, including temperature change rate, hot spot spatial distribution, and abnormal threshold detection. In step S2042, it is necessary to explain in detail that in the construction of the regional health profile, temperature data of all devices in the region are first collected through an IoT sensor network, including device surface temperature, ambient temperature, and temperature rise rate of key nodes. This data, after preprocessing, is input into a thermal imaging analysis model. This model uses a combination of three-dimensional heat conduction equations and infrared radiation theory to construct a regional thermal field distribution map. Using the spatial location information of devices provided by a dynamic digital twin, the temperature data is mapped onto a three-dimensional geometric model, forming a virtual thermal field that completely corresponds to the physical region. A Kalman filter algorithm is used to perform spatiotemporal alignment of the temperature data from multiple devices, eliminating errors caused by sensor sampling delays. Simultaneously, wavelet transform is used to decompose the temperature signal and extract and suppress high-frequency noise components. In the feature extraction stage, the model automatically identifies hotspot areas where the temperature change rate exceeds a threshold (e.g., 5℃ / min) and calculates the spatial distribution density of hotspots (the number of hotspots per unit area). To detect abnormal temperature rises, an anomaly detection module based on a deep belief network is introduced. This module establishes a normal temperature rise pattern library through unsupervised learning. When real-time data deviates from the pattern library by more than three standard deviations, an early warning is triggered. The final generated hot spot evolution trend includes three core indicators: temperature change rate curve (showing the overall temperature rise rate of the area), hot spot spatial distribution heat map (red represents high-risk areas and blue represents normal areas), and abnormal threshold detection results (marking the device number and location that exceeds the safe range).

[0023] In this embodiment, the thermal imaging analysis model is constructed as follows: First, a large amount of temperature data of power distribution network equipment under different operating conditions is collected, including temperature samples under normal and fault conditions. The samples need to cover various equipment types (such as transformers, switchgear, cable joints, etc.) and different load levels and ambient temperature conditions. The collected raw temperature data is preprocessed, including data cleaning to remove outliers and noise interference, linear interpolation to fill missing data, and standardization to map the data to a unified numerical range (such as the [0,1] interval). Subsequently, feature engineering is performed to extract key features related to hot spot evolution, including the transient rate of change of temperature (first derivative), temperature acceleration (second derivative), spatial temperature gradient (temperature difference between adjacent equipment), and time series features (such as autocorrelation coefficient, Fourier transform frequency domain features). The preprocessed data is divided into training set, validation set, and test set. A convolutional neural network (CNN) is used to construct the basic framework of the thermal imaging analysis model, where convolutional layers are used to extract spatial features of the temperature field, pooling layers reduce data dimensionality, and fully connected layers realize feature fusion and classification. To enhance the model's ability to process time-series data, a Long Short-Term Memory (LSTM) network unit is embedded within the CNN framework, forming a CNN-LSTM hybrid model. This model can simultaneously capture the spatial distribution patterns and temporal evolution of temperature data. During model training, the Adam algorithm is used to dynamically adjust the learning rate, and the cross-entropy loss function is used to measure the deviation between the predicted results and the true labels. The network weights are updated using the backpropagation algorithm. To prevent overfitting, a Dropout layer is introduced to randomly deactivate some neurons, and an L2 regularization term is added to the loss function. After multiple rounds of iterative training, the model performance is evaluated on the validation set, using the root mean square error (RMSE) and coefficient of determination (R²) as evaluation metrics. Training stops when the RMSE is below a threshold (e.g., 0.5℃) and R² exceeds 0.95. The final deployed model can receive real-time temperature data input and output regional thermal field distribution maps, temperature change rate curves, and abnormal hotspot detection results. Its hotspot localization accuracy reaches device-level (error ≤ 0.2 meters).

[0024] S2043. At the system level, mechanical fatigue level and insulation performance degradation curve are evaluated by mechanical vibration spectrum analysis and insulation resistance monitoring data, and aging trend is predicted by combining historical data. In step S2043, it is necessary to explain in detail that during the construction of the system-level health profile, mechanical vibration signals are first continuously collected through a vibration sensor network deployed on key equipment (such as transformers), with the sampling frequency set to 20kHz to meet the requirements for capturing high-frequency components. Simultaneously, the insulation resistance value of the equipment is measured periodically (e.g., daily) using an insulation resistance tester, forming a time-series data stream. The vibration signals are preprocessed, including bandpass filtering (with a cutoff frequency set to 0.5-1000Hz) to eliminate environmental noise, and a short-time Fourier transform is used to convert the time-domain signal into a spectrum, extracting the amplitude and energy distribution of characteristic frequency components (such as the equipment's natural frequencies of 100Hz, 200Hz, etc.). For the insulation resistance data, a logarithmic transform is performed to compress the data dynamic range, and the resistance change rate between adjacent measurement periods is calculated. In the feature extraction stage, mechanical fatigue assessment uses a convolutional neural network (CNN) to process vibration spectrum maps. This network contains three convolutional layers (kernel sizes of 3x3, 5x5, and 7x7, respectively) and two fully connected layers. Supervised learning is used to establish a mapping relationship between spectral features and mechanical fatigue levels (divided into five levels: Level I <20% of rated life, Level II 20-40%, Level III 40-60%, Level IV 60-80%, and Level V >80%). The insulation performance degradation analysis module uses a long short-term memory network (LSTM) to process insulation resistance time-series data. The network structure contains two LSTM layers (128 neurons per layer) and one fully connected layer. Unsupervised learning is used to capture the trend of resistance values ​​over time and predict the insulation performance degradation curve for the next 30 days. To improve prediction accuracy, an attention mechanism is introduced to assign higher weights to key time points in historical data (such as rainy seasons and high-temperature periods). The final mechanical fatigue rating is displayed as a 3D bar chart (X-axis for equipment number, Y-axis for fatigue rating, and Z-axis for equipment quantity), while the insulation degradation curve is displayed using a dual-coordinate system (primary coordinate shows resistance changes, secondary coordinate shows degradation rate). Simultaneously, an equipment aging trend prediction model is established, using the mechanical fatigue rating and insulation degradation curve as inputs. The Support Vector Regression (SVR) algorithm is used to predict the remaining useful life (RUL) of the equipment. When the predicted value falls below a threshold (e.g., RUL < 1 year), a system-level early warning is triggered, and the location of high-risk equipment and its expected failure time are marked through a global view of the digital twin.

[0025] S2044 integrates device-level, regional-level, and system-level outputs to generate a multi-dimensional health profile.

[0026] In step S2044, it is necessary to explain in detail that the fusion process employs a multi-source data fusion engine. This engine is built on a distributed computing framework and can process data streams at three levels: device-level, region-level, and system-level, in parallel. First, the electrical health score (HELEC) at the device level, the hotspot evolution trend at the region level (including temperature change rate curves, hotspot spatial distribution heatmaps, and anomaly threshold detection results), and the mechanical fatigue level (3D bar chart) at the system level are spatiotemporally aligned with the insulation performance degradation curve (dual coordinate system graph). Data association is achieved through a unified timestamp and spatial coordinate system provided by a dynamic digital twin. In the feature fusion stage, a weighted fusion algorithm is used to synthesize key indicators at different levels. The weight of device-level indicators is set to 0.4 (focusing on the status of individual devices), the weight of region-level indicators is 0.3 (reflecting the impact of the local environment), and the weight of system-level indicators is 0.3 (reflecting the overall operating trend). Specifically, electrical health scores, hotspot temperature change rates, mechanical fatigue levels (converted to 0-100% values), and insulation performance degradation rates are used as input features. After dimensionality reduction via Principal Component Analysis (PCA), these features are input into a Support Vector Machine (SVM) classifier to generate comprehensive health status labels (five levels: Excellent, Good, Moderate, Poor, and Critical). Simultaneously, a nonlinear mapping model is constructed using a Deep Neural Network (DNN) to fuse multi-source features into a three-dimensional health profile vector. The three dimensions correspond to equipment reliability (0-1 range), regional risk level (0-100 points), and system stability (0-10 levels), respectively. The final result is a multi-dimensional health profile.

[0027] As an optional embodiment of the present invention, optionally, training the fault diagnosis model using a longitudinal federated learning architecture in step S3 includes: S301. Set up a vertical federated learning architecture, including a central coordination node and multiple data holding nodes, where the data holding nodes correspond to different regional equipment of the distribution network, and the central coordination node is used for the aggregation of model parameters of the fault diagnosis model. In step S301, it is necessary to explain in detail that the construction of the vertical federated learning architecture aims to solve the problems of data dispersion and high privacy protection requirements of power distribution network equipment. This architecture adopts a layered design. The central coordination node is deployed in the power grid dispatch center, equipped with high-performance computing servers (such as a cluster equipped with NVIDIA A100 GPUs), responsible for the aggregation and optimization of global model parameters. Data holding nodes are distributed in various regional substations or switching stations, using edge computing devices (such as industrial-grade embedded controllers) for local data preprocessing and model training. To ensure data security, model gradient information is transmitted between nodes through an encrypted channel (using the AES-256 algorithm). The central coordination node only receives encrypted intermediate parameters and cannot directly access the original data. During the node initialization phase, the central coordination node distributes a unified base model (such as a fault feature extraction network based on ResNet50) to all data holding nodes. This model contains trainable convolutional layers and batch normalization layers with fixed parameters. Data-holding nodes load corresponding feature mapping tables based on local device types (e.g., transformers, circuit breakers, cables, etc.), converting raw sensor data (e.g., vibration signals, partial discharge pulses, temperature sequences) into standardized feature vectors (uniform dimension 512×1). To address the issue of non-independent and identically distributed data, the central coordination node periodically (e.g., hourly) pushes global statistics (e.g., feature mean, variance) to each node. Data-holding nodes use this data for local data calibration to ensure consistent data distribution across different regions. During model training, each node employs a synchronous update strategy. After completing each local training cycle (e.g., 100 batches), the gradient information is encrypted and uploaded to the central coordination node. The central coordination node calculates the global gradient average using a secure aggregation algorithm (e.g., SecureAggregation based on homomorphic encryption) and weights and fuses the contributions of each node (weights are dynamically adjusted based on the amount of data per node). To prevent malicious node attacks, a node authentication mechanism based on zero-knowledge proofs is introduced to ensure that only authorized nodes can participate in model training. The final trained fault diagnosis model employs federated transfer learning technology, enabling cross-regional knowledge transfer and model optimization without sharing the original data.

[0028] S302. Define a feature space segmentation scheme based on multi-dimensional health profile data. Each data holding node holds a subset of features from electrical health score, hot spot evolution trend, mechanical fatigue level, and insulation performance degradation curve, which serve as the local feature subset of the data holding node. In step S302, it is necessary to explain in detail that the design of the feature space segmentation scheme must balance data privacy and model training efficiency, while ensuring that the contribution of each node's feature subset to the fault diagnosis task is balanced. Firstly, based on the data hierarchy and physical meaning of the multi-dimensional health profile, the features are divided into four categories: electrical features (including electrical health scores and their historical trends), thermal features (temperature change rate curves, hotspot spatial distribution heatmaps, and abnormal threshold detection results), mechanical features (three-dimensional histograms of mechanical fatigue levels, key frequency components of the vibration spectrum), and insulation features (insulation performance degradation curves, resistance change rates). Considering the distribution characteristics of distribution network equipment, a vertical segmentation strategy based on equipment type is adopted. For example, transformer nodes possess both electrical and thermal features (because transformer faults are often accompanied by temperature anomalies), circuit breaker nodes are assigned both mechanical and electrical features (circuit breaker opening and closing actions are strongly correlated with mechanical fatigue), and cable nodes acquire both thermal and insulation features (cable aging is mainly manifested as decreased insulation performance and localized overheating). To avoid feature redundancy, similar features are further subdivided. For example, thermal features are split into dynamic features (temperature change rate, temperature rise rate) and static features (hotspot spatial distribution density, abnormal equipment number), and mechanical features are decomposed into frequency domain features (natural frequency amplitude) and time domain features (fatigue level evolution trend). The local feature subset ultimately held by each data-holding node must meet two conditions: first, the number of features must be balanced (the difference in feature dimensions between nodes does not exceed 20%), and second, the feature correlation must be below a threshold (selected using Pearson correlation coefficient, retaining feature combinations with a correlation <0.7). For example, a substation in a certain region, acting as a data-holding node, might have a local feature subset including: transformer electrical health score (equipment level), cable joint temperature change rate (regional level), circuit breaker mechanical fatigue level (system level), and its own insulation resistance change rate (system level), forming a hybrid feature set across levels and equipment. To verify the rationality of the segmentation scheme, the information gain ratio is used to evaluate the contribution of each feature subset to the fault diagnosis task, ensuring that each node's local feature subset can provide effective information during model training under the federated learning framework, avoiding model bias due to missing features. The final segmentation scheme is deployed to each data-holding node in the form of a configuration file. The nodes dynamically load local feature subsets according to the configuration file and interact with the central coordination node through a vertical federated learning architecture to achieve a privacy-preserving training mode of "data not leaving the domain and knowledge being transferred together".

[0029] S303. The central coordination node initializes the global parameters of the global fault diagnosis model in the vertical federated learning architecture. The global fault diagnosis model includes convolutional neural network layers and long short-term memory network units. In step S303, it is necessary to explain in detail that the initialization of the global fault diagnosis model is to build a basic model framework with powerful feature extraction and temporal analysis capabilities. First, the global model adopts a hybrid neural network structure, where the Convolutional Neural Network (CNN) layer is responsible for extracting device state features from the spatial dimension, and the Long Short-Term Memory (LSTM) unit is used to capture long-term dependencies in the temporal data. Specifically, the CNN layer contains three convolutional modules, each consisting of a convolutional layer (kernel sizes of 5×5, 3×3, and 1×1 respectively), a batch normalization layer, and a ReLU activation function, extracting feature maps from coarse to fine granular by progressively reducing the receptive field. To enhance the model's adaptability to multi-scale features, a Spatial Pyramid Pooling (SPP) layer is introduced after the second convolutional module to unify feature maps of different scales into a fixed dimension (e.g., 256×1). The LSTM unit adopts a bidirectional (Bi-LSTM) structure, containing 128 hidden units, capable of simultaneously utilizing forward and backward temporal information. Its output is weighted and fused through an attention mechanism to highlight the feature contributions of key time steps. During model initialization, the central coordinating node uses the Xavier initialization method to assign random initial weights to the convolutional kernels of the CNN layers, ensuring that the variance of the input signal remains stable during propagation. For the weight matrix of the LSTM unit, orthogonal initialization is used to alleviate the gradient vanishing problem. To accelerate model convergence, the bias terms of the global model are initialized to 0.1, and the input data is preprocessed using global statistics (such as feature mean and variance). Furthermore, to improve the model's generalization ability to different device types, a device type embedding layer is introduced after the first convolutional module of the CNN layer. This layer encodes the device category into a learnable low-dimensional vector (e.g., 16-dimensional) and concatenates it with the feature map. The final generated global model parameter file (containing the weight matrix, bias terms, and network structure configuration) is distributed to each data-holding node through an encrypted channel as the initial state for local training, ensuring that all nodes start collaboratively from the same starting point under the federated learning framework.

[0030] S304. The data holding node uses a subset of local features to calculate the gradient of the local fault diagnosis model in the vertical federated learning architecture, protects privacy through homomorphic encryption, and transmits the encrypted gradient to the central coordination node. In step S304, it is necessary to explain in detail that after receiving the global model parameters distributed by the central coordination node, the data holding node first inputs a local feature subset into the local fault diagnosis model (this model has the same structure as the global model, but the parameters are copies of the initialized global parameters). During the forward propagation of the model, the CNN layer performs convolution operations on the spatial features (such as the two-dimensional trend map of electrical health score and the heat map of hot spot spatial distribution) in the local feature subset, extracts local patterns (such as edges and textures) through learnable convolution kernels, and generates multi-channel feature maps; subsequently, the batch normalization layer standardizes the feature maps, eliminates internal covariate bias, and accelerates model convergence; the ReLU activation function introduces nonlinear transformation to enhance the model's expressive power. For temporal features (such as the evolution trend of mechanical fatigue level and the insulation resistance change rate curve), bidirectional temporal modeling is performed by Bi-LSTM units. The forward LSTM captures the impact of historical states on the current state, and the backward LSTM mines the feedback of future information on the current state. Its output is dynamically weighted through an attention mechanism to highlight the feature contributions of key time steps (such as the fault precursor stage). After forward propagation, the model outputs the predicted results (such as the probability distribution of fault types), and calculates the loss function value (such as cross-entropy loss) against the locally labeled ground truth labels (if partially labeled data exists). During backpropagation, the gradient of the loss function with respect to the model parameters is calculated layer by layer according to the chain rule. The gradient of the CNN layer is propagated from the output layer to the input layer through the backpropagation algorithm, updating the convolutional kernel weights; the gradient of the LSTM unit is calculated backward along the time step using the Backpropagation-Time Trial (BPTT) algorithm, adjusting the connection weights between hidden units. To protect data privacy, the data holding node uses homomorphic encryption technology (such as the Paillier encryption algorithm) to encrypt the gradient information, mapping the original gradient to a ciphertext space to generate an encrypted gradient packet. The encrypted gradient packet is transmitted to the central coordination node through an encrypted channel (such as the TLS / SSL protocol). During transmission, digital signature technology (such as the RSA algorithm) is used to verify data integrity and prevent gradient tampering.

[0031] S305. The central coordination node aggregates all encryption gradients, applies a weighted average algorithm to update the global model parameters of the global fault diagnosis model, and broadcasts the updated global model parameters to each data holding node. In step S305, it is necessary to explain in detail that after receiving the encrypted gradient packets uploaded by each data-holding node, the central coordination node first decrypts the ciphertext using homomorphic decryption technology (corresponding to the encryption algorithm used by the data-holding node) to recover the original gradient information. To address the problem of uneven gradient contributions caused by differences in the amount of data among different nodes, a weighted average algorithm is used to aggregate the gradients. The weight coefficients are dynamically calculated based on the amount of local feature subset data held by each node (for example, nodes with larger data volumes have higher weights, but normalization is required to ensure that all weights sum to 1). Specifically, for each model parameter (such as a certain weight value of a CNN convolutional kernel), the central coordination node collects the gradient values ​​of all nodes for that parameter, multiplies them by the corresponding weight coefficients, sums them, and then divides by the total weight sum to obtain the global gradient average value of that parameter. To prevent individual nodes from generating abnormal gradients due to data quality issues (such as noise interference or missing features), a gradient pruning mechanism is introduced to replace gradient values ​​that exceed a preset range (such as mean ± 3 standard deviations) with boundary values. After gradient aggregation is completed, the central coordinating node uses the stochastic gradient descent (SGD) optimization algorithm (or its variant such as Adam) to update the global model parameters. The update step size (learning rate) is dynamically adjusted according to the model training stage (larger in the early stage to accelerate convergence, and smaller in the later stage for fine optimization).

[0032] S306. Iteratively execute steps S304-S305, detect the training termination condition through the convergence of the loss function, and output the trained fault diagnosis model when the accuracy of the model on the validation set exceeds the preset threshold.

[0033] In step S306, it needs to be specifically explained that the goal of the iterative training process is to gradually improve the performance of the global fault diagnosis model through multiple rounds of interaction between the data holding nodes and the central coordination node. Specifically, in each iteration, the data holding node first calculates the local gradient based on the local feature subset and the current global model parameters. Then, the central coordination node aggregates all encrypted gradients and updates the global model parameters. To ensure the effectiveness of the training process and the rationality of the termination condition, a loss function convergence detection mechanism and a validation set evaluation strategy need to be introduced. First, the choice of loss function needs to match the fault diagnosis task type: for classification tasks (such as fault type identification), the cross-entropy loss function is used to measure the difference between the predicted probability distribution and the true label; for regression tasks (such as remaining life prediction), the mean squared error (MSE) or mean absolute error (MAE) is used to quantify the deviation between the predicted value and the true value. After each iteration, the central coordination node calculates the loss value of the global model on the validation set and records its changing trend. If the loss value decreases by less than a preset threshold (e.g., 0.001) for multiple consecutive rounds (e.g., 5 rounds), or if the loss value fluctuates (e.g., the standard deviation exceeds 0.01), the loss function is considered to have converged, and the training process can be terminated early. Simultaneously, to avoid overfitting, the model's accuracy, recall, F1 score, and other metrics on the validation set must be monitored. When the accuracy exceeds a preset threshold (e.g., 95%) and remains stable (e.g., an improvement of no more than 0.5% for 3 consecutive rounds), the current global model parameters are output as the completed fault diagnosis model. The final generated fault diagnosis model is stored in a standardized format (e.g., ONNX), including model structure definitions, weight parameters, and preprocessing configurations. It can be directly deployed to edge devices in the power distribution network or cloud servers to achieve real-time fault detection and early warning.

[0034] As an optional embodiment of the present invention, optionally, extracting fault features based on a fault diagnosis model using a multi-dimensional health profile in step S4 includes: S401. Preprocess the multi-dimensional health profile to obtain the preprocessed feature matrix; In step S401, it is necessary to explain in detail that the preprocessing of the multi-dimensional health profile involves transforming the raw data into a structured and standardized feature matrix. First, for different types of data sources (such as real-time sensor monitoring data, equipment historical maintenance records, inspection images, etc.), a categorized preprocessing strategy is adopted: For numerical data (such as electrical health scores, temperature change rates), missing value imputation (using mean or median interpolation) and outlier processing (removing outliers based on the 3σ principle) are performed, and then the data is mapped to the [0,1] interval through min-max normalization to eliminate dimensional differences; For categorical data (such as equipment type, fault type labels), one-hot encoding is used to convert it into binary vectors to avoid biases introduced by manual assignment; For time-series data (such as mechanical fatigue level evolution curves, insulation resistance change rates), a fixed-length time-series segment is extracted using the sliding window method (the window length is dynamically adjusted according to equipment characteristics, such as 72 hours for transformers and 24 hours for cables), and linear interpolation is used to unify the time step to ensure the continuity of the time-series data. Secondly, to address the cross-level characteristics (equipment level, regional level, system level) of multi-dimensional health profiles, a hierarchical feature extraction framework is constructed: equipment-level data (such as the electrical health score of a single transformer) is directly used as the basic feature; regional-level data (such as the average temperature of all cable joints in a substation) generates regional features through spatial aggregation (such as weighted averaging); system-level data (such as the load fluctuation rate of the entire distribution network) extracts system features through global statistics (such as percentile calculation). To enhance the distinguishability between features, derivative transformations are performed on similar features: for example, the first derivative (temperature rise acceleration) and second derivative (temperature rise curvature) of the temperature change rate are calculated to capture dynamic trends; statistical features (such as maximum value, minimum value, variance) and frequency domain features (obtaining the dominant frequency component through Fast Fourier Transform (FFT)) are extracted from the mechanical fatigue level. Finally, all processed features are matrix-concatenated according to equipment type, data level, and feature type (numerical, categorical, time-series) to generate a feature matrix of dimension .

[0035] S402. The convolutional neural network layer of the fault diagnosis model is used to extract the local features of the hot spot spatial distribution features in the feature matrix by convolution kernel, and the hot spot spatial feature vector is output after the dimension is compressed by the max pooling layer. In step S402, it is necessary to explain in detail that, firstly, the convolutional layer slides across the hotspot feature map using preset convolutional kernels. Each convolutional kernel acts as a feature detector, capturing patterns in local regions (such as edges, textures, and shapes). For example, small-sized convolutional kernels (3×3) are good at extracting fine-grained features (such as gradient changes at the edges of hotspots), while large-sized convolutional kernels (5×5) can capture coarser-grained spatial structures (such as hotspot clustering areas). Through multi-channel convolution operations, the model generates multiple sets of feature maps, each corresponding to a specific pattern extracted by a convolutional kernel. Subsequently, a batch normalization layer standardizes the feature maps, normalizing the feature values ​​of each channel to a distribution with a mean of 0 and a variance of 1, alleviating the internal covariate shift problem and accelerating model convergence. The ReLU activation function introduces a non-linear transformation, setting negative values ​​to zero, retaining positive values, and enhancing feature expressiveness, enabling the model to fit complex spatial relationships. To reduce feature dimensionality while preserving key information, the max pooling layer slides across the feature map with a fixed stride (e.g., 2×2), selecting the maximum value of each local region as the output to generate a compressed hotspot spatial feature vector. For example, if the original feature map size is 64×64, after 2×2 max pooling, the output size becomes 32×32, reducing the number of features by 75% while retaining the most salient spatial patterns. If the feature map size is not an integer multiple of the pooling stride, boundary padding (e.g., zero padding) ensures size matching. The final generated hotspot spatial feature vector will be fused with other types of features (e.g., temporal features, statistical features).

[0036] S403. Use the long short-term memory network unit of the fault diagnosis model to process the electrical health score time series data in the feature matrix, capture the time-series dependencies and output the electrical time series feature vector. In step S403, it is necessary to explain in detail that when processing electrical health score time series data, the Long Short-Term Memory (LSTM) network unit dynamically controls the flow of information through its unique gating mechanism (input gate, forget gate, output gate). Specifically, the input gate determines how much new information is added to the cell state at the current time step, the forget gate determines how much of the cell state from the previous time step is retained at the current time step, and the output gate controls how much of the current cell state is output to the hidden state. For the electrical health score time series, the score value at each time step is first used as input along with the hidden state from the previous time step, and the input gate filters out the information that needs to be updated; at the same time, the forget gate determines which old information needs to be forgotten based on the current input and the previous hidden state; after the cell state is updated, the output gate generates the hidden state of the current time step, which not only contains the information of the current time step but also implicitly contains the dependencies of all previous time steps. To capture bidirectional temporal dependencies (i.e., the influence of the past on the present and the feedback of the future on the present), a bidirectional LSTM (Bi-LSTM) structure is adopted. The forward LSTM and backward LSTM process the forward and reverse directions of the sequence, respectively, and their outputs are concatenated at the last time step to form an electrical temporal feature vector that integrates bidirectional information. For example, for an electrical health rating sequence of length T, the forward LSTM processes the sequence from t=1 to t=T, and the backward LSTM processes it from t=T to t=1. Finally, at time t=T, the forward and backward hidden states are concatenated to obtain a feature vector of dimension 2×H (H is the dimension of a single LSTM hidden unit). To further highlight the feature contributions of key time steps (such as the stage of abnormal rating fluctuations), an attention mechanism can be introduced after the Bi-LSTM output. By calculating the weights of the hidden states at each time step (e.g., based on a soft attention mechanism, the weights are related to the similarity between the hidden state and the global context), the hidden states are weighted and summed to generate the final electrical temporal feature vector.

[0037] S404. A multi-head self-attention mechanism is used to dynamically assign weights to the mechanical fatigue level and insulation performance degradation curve to generate an equipment-level degradation feature vector. In step S404, it is necessary to explain in detail that the multi-head self-attention mechanism captures the dynamic correlation between mechanical fatigue level and insulation performance degradation curve from different subspaces by computing multiple attention heads in parallel. Specifically, each attention head first maps the input sequence (such as the concatenated vector of time-series data of mechanical fatigue level and insulation performance degradation curve) into a query vector (Q), a key vector (K), and a value vector (V), where Q is used to query relevant information, K is used to match information, and V is used to extract information. For example, for a sequence of length T, the dimensions of Q, K, and V are all T×dk (dk is the dimension of the key vector). Subsequently, the weight of each time step is calculated by scaling dot product attention: multiplying Q by the transpose of K and dividing by the square root of dk yields the attention score matrix, which is then normalized to a weight matrix using the Softmax function. This matrix represents the degree of dependence of each time step on other time steps. For example, if the mechanical fatigue level of a certain time step shows an abnormal increase, the attention value of that time step on adjacent time steps in its corresponding weight matrix will increase significantly, highlighting the local acceleration characteristics of the degradation process. To capture multi-dimensional correlations, the multi-head mechanism divides the input sequence into multiple subspaces (e.g., 8 heads). Each head independently calculates attention weights and then concatenates them to generate intermediate features of dimension T×(8×dv) (dv is the dimension of the value vector). Finally, the intermediate features are compressed into a device-level degradation feature vector through linear transformation. This vector integrates the co-degradation patterns of mechanical fatigue and insulation performance. For example, if the insulation performance degradation curve accelerates after the mechanical fatigue level reaches a threshold, the feature vector will strengthen this temporal coupling relationship through weight allocation.

[0038] S405. In the data holding node of the vertical federated learning architecture, the hot spot spatial feature vector, electrical time series feature vector and device-level degradation feature vector are spliced ​​and fused, and the attention mechanism is used to assign different weights to different types of feature vectors to generate a fused feature vector. In step S405, it is necessary to explain in detail that, under the vertical federated learning architecture, the data holding node first concatenates the hotspot spatial feature vector, the electrical temporal feature vector, and the device-level degradation feature vector to form a preliminary fused feature matrix. To handle the semantic differences between different types of features (spatial, temporal, and degradation), an attention mechanism is used to dynamically allocate weights: each feature vector is transformed linearly to generate a query vector (Q), a key vector (K), and a value vector (V), where Q is used for query relevance, K for matching information, and V for feature extraction. For example, the Q of the hotspot spatial feature vector may be multiplied by the K of the electrical temporal feature vector to obtain the correlation strength between the two. Weights are calculated by scaling the dot product attention: multiplying Q by the transpose of K and dividing by the square root of the feature dimension yields the attention score matrix, which is then normalized to a weight matrix using the Softmax function. This matrix represents the contribution of different types of features to the final fusion; for example, if the electrical temporal feature is more critical in fault diagnosis, its corresponding weight value will be significantly higher than other features. To capture multi-dimensional correlations, a multi-head attention mechanism is employed, dividing the input features into multiple subspaces (e.g., four heads). Each head independently calculates its attention weights, and the results are concatenated to generate intermediate features of dimension 1. Finally, a linear transformation compresses these intermediate features into a fused feature vector. This vector preserves local details of the original features (e.g., edge gradients of hotspots) while incorporating collaborative patterns across different feature types (e.g., the temporal coupling between electrical rating anomalies and mechanical fatigue acceleration). To enhance feature robustness, residual connections are introduced, adding the original concatenated features to the attention output to avoid the vanishing gradient problem. Simultaneously, layer normalization is used to standardize the fused features, normalizing the mean to 0 and the variance to 1 for each feature, mitigating internal covariate bias and accelerating model convergence. The final fused feature vector will serve as the input to the fault diagnosis model.

[0039] S406. Input the fused feature vector into the fully connected layer of the fault diagnosis model. The fully connected layer performs nonlinear transformation and feature mapping on the fused features, and finally outputs the fault feature vector.

[0040] In step S406, it needs to be explained in detail that the fully connected layer, as a fault diagnosis model, functions by performing a high-dimensional nonlinear mapping on the fused feature vector to extract more discriminative fault features. Specifically, the fully connected layer consists of multiple neurons (nodes). Each neuron performs a linear transformation on the input features through a weight matrix and a bias vector, and then introduces nonlinear factors through an activation function (such as ReLU, Sigmoid, or Tanh) to enhance the model's expressive power. For example, if the dimension of the fused feature vector is D and the fully connected layer contains N neurons, then the dimension of the weight matrix is ​​D×N, and the dimension of the bias vector is N. The input features are multiplied by the weight matrix and then added to the bias vector to obtain the original output of each neuron, which is then converted into nonlinear features through an activation function. To prevent overfitting, a Dropout mechanism can be introduced after the fully connected layer to randomly disable some neurons (e.g., with a masking probability of 0.5), forcing the model to learn a more robust feature representation. Furthermore, to accelerate model convergence and alleviate the gradient vanishing problem, batch normalization can be used to standardize the output of each layer, adjusting the feature distribution to a distribution with a mean of 0 and a variance of 1. Finally, the output of the fully connected layer is a fault feature vector, the dimension of which is determined by the number of neurons in the last layer. This vector is directly used for subsequent fault classification or regression tasks (such as fault type identification or remaining lifetime prediction). For example, to identify 5 types of faults, the last layer can have 5 neurons and use the Softmax activation function to output the probability distribution of each fault category; if remaining lifetime prediction is required, a linear activation function can be used to output continuous values. Through the nonlinear transformation of the fully connected layer, the model can extract higher-level semantic information from the fused features. For example, by combining the spatial distribution of hot spots, electrical temporal variations, and equipment degradation patterns, it can identify local overheating faults caused by insulation aging or structural loosening faults caused by mechanical fatigue accumulation.

[0041] As an optional embodiment of the present invention, optionally, in step S5, a three-level output model is constructed based on a bio-inspired swarm intelligence optimization algorithm and a deep Q-network to perform end-to-end optimization on the fault feature vector, generating a diagnostic result that includes assessments of fault type, location, and severity. S501. Construct a three-level output model based on fault feature vectors. The three-level output model includes a feature fusion layer, a decision optimization layer, and a diagnostic output layer. The feature fusion layer uses a gated recurrent unit network to fuse multi-source fault features. In step S501, it is necessary to explain in detail that the three-level output model takes the fault feature vector as input and achieves end-to-end fault diagnosis optimization through a hierarchical architecture. The feature fusion layer uses a gated recurrent unit (GRU) network to process multi-source heterogeneous features. Its core mechanism dynamically regulates information flow through the reset gate and update gate. Specifically, the reset gate calculates the reset probability based on the current input and the hidden state of the previous time step, deciding whether to discard parts of historical information that are irrelevant to the current task (such as redundant features under normal operating conditions). The update gate generates a retention probability through the sigmoid function, controlling the fusion ratio of new information and historical memory. For example, when an abnormal fluctuation is detected in the electrical timing features, the update gate increases the weight of that time step, strengthening the transmission of fault-related features. The GRU network gradually extracts high-level semantic features by stacking multiple layers (such as 3 layers), with the output dimension of each layer decreasing layer by layer (such as from 256 dimensions to 64 dimensions), finally generating a fused global feature vector.

[0042] S502. Initialize the policy network parameters of the deep Q network using a bio-inspired swarm intelligence optimization algorithm. The bio-inspired swarm intelligence optimization algorithm adopts the particle swarm optimization algorithm, which searches for the optimal policy weight in the solution space by simulating swarm intelligence behavior. In step S502, it is necessary to explain in detail that the Particle Swarm Optimization (PSO) algorithm finds the optimal solution by simulating the swarm behavior of birds or fish. During the parameter initialization phase of the Deep Q-Network (DQN), the PSO algorithm is used to search for the optimal policy network weights to improve the initial performance and convergence speed of the model. Specifically, the PSO algorithm treats each potential policy network weight combination as a particle. These particles dynamically adjust their positions (i.e., weight values) in the solution space based on their own historical best position (individual extremum) and the swarm's historical best position (global extremum). Each particle is iteratively optimized using velocity update formulas and position update formulas. The velocity update formula comprehensively considers the particle's current velocity, the difference between the individual extremum and the current position, the difference between the global extremum and the current position, and a random perturbation term to enhance exploration capability; the position update formula adjusts the particle's position in the solution space based on the updated velocity. For example, when initializing the DQN policy network, assuming the solution space is a 100-dimensional weight vector, the PSO algorithm will generate a swarm containing 50 particles, each representing a 100-dimensional weight combination. Through multiple iterations (e.g., 100 times), particles gradually converge towards the global optimum, and the weights corresponding to the globally optimal particles are ultimately selected as the initial parameters of the DQN policy network. This initialization method effectively avoids the local optimum problem that may be caused by random initialization, improving the training efficiency and diagnostic accuracy of the model.

[0043] S503. Set up a reinforcement learning environment in a deep Q network, take the fault feature vector as the state input, define the fault diagnosis action space including three types of actions: fault type classification, location localization and severity quantification, and construct a reward function based on historical fault data. The expression for the reward function is: , , , ; , , ; in, The total reward value for the entire fault diagnosis task is an indicator that comprehensively considers three aspects: fault type classification, location, and severity quantification. The dynamic weights represent the fault type classification task; they are a subset of the weights for the current training round. The function is used to adjust the proportion of fault type classification rewards in the total reward. The reward value for the fault type classification task is determined based on whether the classification is correct and the confidence level of the classification. The dynamic weights representing the location localization task are also related to the current training round. This function is used to adjust the proportion of location-based rewards in the total rewards. The reward value for the location positioning task is based on the distance between the predicted and actual locations. To determine, The dynamic weights representing the severity of the task are also related to the current training epoch. The function is used to adjust the weight of severity-quantified rewards in the total reward. The reward value represents the severity quantification of the task, based on the error between the predicted severity score and the actual severity score. To determine, Indicates the total number of training rounds. Indicates the current training round. This indicates the number of different fault types that may occur during a fault diagnosis task. Represents the natural base. This represents the distance between the predicted location and the actual location, used to measure the accuracy of location positioning. This represents the maximum acceptable distance threshold, a pre-set value. When the distance between the predicted and actual locations exceeds this threshold, the location is considered severely inaccurate. This represents the error between the predicted severity score and the actual severity score, used to measure the accuracy of severity quantification. This represents the maximum permissible error, which is a pre-set value. When the error exceeds this threshold, the severity quantification is considered severely inaccurate.

[0044] In step S503, it is necessary to explain in detail that in the deep Q-network, the reinforcement learning environment uses the fault feature vector as the state input. This vector integrates multi-source information such as hotspot spatial distribution, electrical temporal changes, and equipment degradation patterns, providing the model with a comprehensive fault representation. To define the fault diagnosis action space, the system designs three core actions: fault type classification, location localization, and severity quantification. The fault type classification action aims to map the input features to a preset set of fault categories (such as insulation aging, mechanical loosening, etc.); the location localization action determines the specific location of the fault (such as a section of the line or a component of the equipment) through regression or classification; the severity quantification action outputs a continuous value or discrete level, representing the severity of the fault (such as mild, moderate, or severe). To guide the model in learning the optimal strategy, a dynamic reward function is constructed based on historical fault data. This function adopts a multi-task weighted summation form, comprehensively considering the performance of the three sub-tasks: fault type classification, location localization, and severity quantification. Specifically, the total reward consists of three parts: fault type classification reward, location positioning reward, and severity quantification reward. Each part is multiplied by a dynamic weighting coefficient, which is adjusted with each training round to balance the learning progress of different tasks. For example, in the early stages of training, the fault type classification task may have a higher weight to quickly converge to the main fault mode; as training progresses, the weights of location positioning and severity quantification tasks gradually increase to improve the precision of diagnosis. The fault type classification reward is determined based on classification accuracy and confidence. If the classification is correct and the confidence is high, a positive reward is given; conversely, if the classification is incorrect or the confidence is low, a negative reward is given. The location positioning reward is calculated based on the geometric distance between the predicted location and the actual location. The closer the distance, the higher the reward. When the distance exceeds a preset threshold, the reward is reset to zero to penalize severely inaccurate positioning. The severity quantification reward is determined based on the error between the predicted score and the actual score. The smaller the error, the higher the reward. When the error exceeds an allowable threshold, a penalty is also imposed. Through this dynamic reward mechanism, the model can gradually optimize the performance of each subtask during training, ultimately achieving accurate diagnosis of fault type, location, and severity.

[0045] S504. The three-level output model is optimized through an end-to-end training process based on a deep Q-network, including: performing dimensionality reduction and normalization on the fault feature vector in the feature fusion layer to generate a fused feature matrix; and calculating the Q value of each action based on the fused feature matrix using a deep Q-network in the decision optimization layer, and selecting the optimal diagnostic action sequence based on an ε-greedy strategy. The expression for calculating the Q-value of each action using a deep Q-network based on the fused feature matrix in the decision optimization layer is as follows: ;in, Indicates the state Take action below At that time, the deep Q-network adjusts according to the current parameters. The calculated Q-value represents an estimate of the expected cumulative reward that can be obtained by taking a certain action in a certain state. Indicates at time step At that time, the fused feature matrix (which can be viewed as a state vector) contains fused feature information obtained from the fault feature vector after processing by the feature fusion layer, and is used to describe the current fault diagnosis state. Indicates at time step The actions taken during diagnosis, within the fault diagnosis action space, include three categories: fault type classification, location identification, and severity quantification. This represents the set of parameters of a deep Q-network, including the weight matrix and bias vector, etc. These parameters are continuously updated and optimized during training to enable the Q-network to estimate the Q-value more accurately. This represents the weight vector of the output layer, used to map the output of the hidden layer to Q-values. This represents the activation function. This represents the weight matrix of the hidden layer, used to weight the input state vector. Perform a linear transformation. This represents the bias vector of the hidden layer and the weight matrix. In conjunction with this, the result of the linear transformation is shifted. This represents the bias scalar of the output layer, used to shift and adjust the final Q-value calculation result; In step S504, it is necessary to explain in detail that the goal of the end-to-end training process is to enable the model to gradually learn the mapping relationship from the fault feature vector to the optimal diagnostic action by iteratively updating the parameters of the deep Q-network. Specifically, the training process includes the following key steps: First, the input fault feature vector is subjected to dimensionality reduction and normalization in the feature fusion layer. Dimensionality reduction is achieved through techniques such as principal component analysis (PCA) or autoencoder, aiming to remove redundant information and retain the most discriminative features, for example, compressing the original 256-dimensional feature vector to 64 dimensions; normalization uses Min-Max normalization or Z-score normalization methods to map the feature values ​​to a fixed range (such as [0,1] or [-1,1]) to eliminate the influence of dimensional differences on model training. The processed feature matrix serves as the input to the decision optimization layer, providing a unified data representation for subsequent Q-value calculation. Second, in the decision optimization layer, the deep Q-network calculates the Q-value of each action based on the fused feature matrix. The Q-value is a core concept in reinforcement learning, representing an estimate of the expected cumulative reward that can be obtained by taking a certain action in a specific state. The computation process is implemented through forward propagation: the input state vector (i.e., the fused feature matrix) first undergoes a linear transformation in the hidden layer (completed by the weight matrix and bias vector), and then nonlinear characteristics are introduced through an activation function (such as ReLU or Tanh). The hidden layer output is then multiplied by the output layer weight vector and the output layer bias scalar is added to finally generate the Q-value for each action. For example, if the fault diagnosis action space contains 3 types of faults, 5 possible locations, and 4 severity levels, the output layer will generate 3+5+4=12 Q-values, each corresponding to a specific action. To balance exploration and exploitation, the model adopts an ε-greedy strategy for action selection: at each decision step, an action is randomly selected with probability ε (exploration), and the action with the largest current Q-value is selected with probability 1-ε (exploitation). The ε-value typically decays dynamically with training rounds, initially higher to encourage exploration, and gradually decreasing later to stabilize the optimal exploitation strategy. Finally, the parameters of the deep Q-network are updated through the backpropagation algorithm. The model calculates a loss function (such as mean squared error or Huber loss) based on reward signals from environmental feedback and adjusts network weights and biases using gradient descent (such as the Adam optimizer) to gradually approximate the predicted Q-value with the actual cumulative reward. For example, if a certain action selection results in a large location error, the reward value for the corresponding location task will be low. During backpropagation, the gradient update magnitude of the Q-value for that action will be increased, prompting the model to tend to choose more accurate actions in similar situations in the future. Through multiple rounds of iterative training, the model can eventually generate high-precision diagnostic results that include assessments of fault type, location, and severity.

[0046] S505. Based on the optimal diagnostic action sequence, the action sequence is parsed at the diagnostic output layer to generate a diagnostic result containing the fault type probability distribution, location coordinate confidence interval, and severity score vector. The severity score vector is based on the low, medium, and high quantized values ​​output by the fuzzy logic evaluation system.

[0047] In step S505, it is necessary to explain in detail that the diagnostic output layer first parses the action sequence, classifies the actions according to fault type, and uses the Softmax function to convert the Q-value of each action into a probability distribution, generating a fault type probability distribution vector. Each element corresponds to the probability of occurrence of a preset fault type, and the sum of all elements is 1. For example, if the preset fault types include 5 categories such as insulation aging, mechanical loosening, and poor contact, then a 5-dimensional vector is output, such as [0.3, 0.25, 0.2, 0.15, 0.1], indicating that insulation aging has the highest probability (30%), mechanical loosening has a probability of 25%, and so on. For location positioning actions, the system uses regression or classification methods to generate confidence intervals for position coordinates. If a regression method is used, the coordinates of the predicted location are directly output (e.g., x=120m, y=50m for section 3 of the line), and the confidence interval is calculated based on the prediction error distribution (e.g., Gaussian distribution) (e.g., the 95% confidence interval is [115m, 125m]). If a classification method is used, the section or component number where the fault may occur is output, and the confidence score for each section is given (e.g., the confidence score for section 3 is 0.8, for section 2 it is 0.15, and for other sections it is 0.05). For the severity quantification action, the system introduces a fuzzy logic evaluation system, which maps continuous prediction scores (e.g., values ​​between 0 and 1) to three discrete levels: low, medium, and high. Specifically, the system pre-defines three fuzzy sets (low, medium, and high), each corresponding to a membership function (such as a triangular or trapezoidal function). The input score is used to calculate the membership degree to each level (e.g., a score of 0.6 has a membership degree of 0.8 for "medium," 0.2 for "high," and 0 for "low"). Finally, the level with the highest membership degree is selected as the severity assessment result (e.g., "medium"). For example, if the predicted severity score is 0.6, the output severity score vector is [0, 0.8, 0.2], indicating that the "medium" level has the highest confidence (80%). Through the above analysis process, the diagnostic output layer finally generates a structured diagnostic result, including the fault type probability distribution (e.g., a 5-dimensional vector), the location coordinate confidence interval (e.g., [115m, 125m]), and the severity score vector (e.g., [0, 0.8, 0.2]).

[0048] Example 2 A 10 kV distribution network fault detection system based on the Internet of Things includes: processor; Memory used to store processor-executable instructions; The processor is configured to implement an IoT-based 10 kV distribution network fault detection method when executing executable instructions.

[0049] It should be noted that the computer device includes a processor, a memory, and may also include one or more of a multimedia component, an input / output (I / O) interface, and a communication component.

[0050] The processor controls the overall operation of the computer device to complete all or part of the steps in the above-mentioned IoT-based 10 kV distribution network fault detection method.

[0051] Memory is used to store various types of data to support the operation of the computer device. This data may include, for example, instructions for any application or method used to operate on the computer device, as well as application-related data. Memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0052] The multimedia component may include a screen and an audio component, wherein the screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals; for example, the audio component may include a microphone for receiving external audio signals, the received audio signals may be further stored in memory or transmitted via a communication component; the audio component may also include at least one speaker for outputting audio signals.

[0053] I / O interfaces provide interfaces between the processor and other interface modules, such as keyboards, mice, buttons, etc.; these buttons can be virtual buttons or physical buttons.

[0054] The communication component is used for wired or wireless communication between the computer device and other devices; wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G or 5G, or one or more combinations thereof, and the corresponding communication component may include: Wi-Fi module, Bluetooth module, NFC module, mobile communication module.

[0055] As a preferred embodiment, the computer device may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the above-described IoT-based 10 kV distribution network fault detection method.

[0056] Although embodiments of the invention have been shown and described, those skilled in the art will understand 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 claims and their equivalents.

Claims

1. A method for detecting faults in a 10 kV power distribution network based on the Internet of Things, characterized by, The method comprises: S1, constructing a power distribution network dynamic digital twin; S2, verifying the equipment health state index based on the dynamic digital twin using a ring verification mechanism, and generating a multi-dimensional health portrait; S3, training a fault diagnosis model using a longitudinal federated learning architecture; S4, extracting fault features based on the multi-dimensional health portrait using the fault diagnosis model; S5, based on a biological heuristic swarm intelligence optimization algorithm and a deep Q network, a three-level output model is constructed to optimize the fault feature vector end-to-end, and a diagnosis result containing fault type, location and severity evaluation is generated.

2. The 10 kV distribution network fault detection method based on the Internet of Things according to claim 1, characterized in that, In step S1, constructing a power distribution network dynamic digital twin comprises: S101, establishing a three-dimensional geometric model based on laser point cloud data; S102, collecting multi-modal data of the power grid, mapping the multi-modal data to the three-dimensional geometric model, and fusing the multi-modal data through Kalman filtering and deep belief network to eliminate noise interference and correct local overheating area positioning error, forming a digital twin basic framework; S103, based on the real-time collected modal data, the digital twin basic framework is updated in real time to obtain a power distribution network dynamic digital twin.

3. The 10 kV distribution network fault detection method based on the Internet of Things according to claim 1, characterized in that, In step S2, based on the dynamic digital twin, the equipment health state index is verified using a ring verification mechanism, and a multi-dimensional health portrait is generated, comprising: S201, based on the dynamic digital twin, multi-source real-time data flow of the power distribution network is obtained, and an electrical-thermal-mechanical-electromagnetic four-dimensional data matrix is constructed; S202, using outer ring Kalman filtering to time align and filter power frequency noise of the electrical-thermal-mechanical-electromagnetic four-dimensional data matrix, and using inner ring deep belief network to fuse features of the electrical-thermal-mechanical-electromagnetic four-dimensional data matrix, to obtain an equipment health state feature vector; S203, updating the dynamic digital twin using the equipment health state feature vector to generate a health state index; S204, based on the health state index, a multi-dimensional health portrait is constructed from three scales of device level, regional level and system level, and the multi-dimensional health portrait includes electrical health score, thermal spot evolution trend, mechanical fatigue level and insulation performance degradation curve.

4. The 10 kV distribution network fault detection method based on the Internet of Things according to claim 3, characterized in that, In step S204, based on the health state index, a multi-dimensional health portrait is constructed from three scales of device level, regional level and system level, comprising: S2041, at the device level, based on the electrical parameters in the health state index, a weighted scoring algorithm is applied to calculate the electrical health score, and the electrical health score is updated in real time; S2042, at the regional level, the health data of multiple devices is integrated, and a thermal imaging analysis model is used to generate a thermal spot evolution trend, including temperature change rate, hot spot spatial distribution and abnormal threshold detection; S2043, at the system level, the mechanical fatigue level and the insulation performance degradation curve are evaluated through mechanical vibration spectrum analysis and insulation resistance monitoring data, and the aging trend is predicted in combination with historical data; S2044, the outputs of the device level, the regional level and the system level are fused to generate a multi-dimensional health portrait.

5. The method for fault detection in 10 kV distribution network based on Internet of Things as claimed in claim 1 wherein, In step S3, a fault diagnosis model is trained using a longitudinal federated learning architecture, comprising: S301, set a longitudinal federal learning architecture, including a central coordination node and a plurality of data holding nodes, wherein the data holding nodes correspond to different regional devices of a power distribution network, and the central coordination node is used for model parameter aggregation of a fault diagnosis model; S302, define a feature space segmentation scheme based on multi-dimensional health profile data, each data holding node holds a part of feature subsets of electrical health score, thermal spot evolution trend, mechanical fatigue level, and insulation performance degradation curve as a local feature subset of the data holding node; S303, the central coordination node initializes global parameters of a global fault diagnosis model in the longitudinal federal learning architecture, and the global fault diagnosis model includes a convolutional neural network layer and a long short-term memory network unit; S304, the data holding node calculates gradients of a local fault diagnosis model in the longitudinal federal learning architecture by using the local feature subset, protects privacy by using a homomorphic encryption technology, and transmits the encrypted gradients to the central coordination node; S305, the central coordination node aggregates all the encrypted gradients, updates global model parameters of the global fault diagnosis model by using a weighted average algorithm, and broadcasts the updated global model parameters to each data holding node; S306, iteratively execute steps S304-S305, detect a training termination condition by loss function convergence, and output a trained fault diagnosis model when the accuracy of the model on a validation set exceeds a preset threshold.

6. The 10 kV distribution network fault detection method based on the Internet of Things according to claim 1 or 5, characterized in that, In step S4, extracting fault features by using the fault diagnosis model based on the multi-dimensional health profile includes: S401, pre-processing the multi-dimensional health profile to obtain a pre-processed feature matrix; S402, using a convolutional neural network layer of the fault diagnosis model to perform kernel local feature extraction on thermal spot spatial distribution features in the feature matrix, and outputting a thermal spot spatial feature vector after dimension compression by a max-pooling layer; S403, using a long short-term memory network unit of the fault diagnosis model to process electrical health score time series data in the feature matrix, capturing time sequence dependence and outputting an electrical time sequence feature vector; S404, using a multi-head self-attention mechanism to dynamically allocate weights for mechanical fatigue level and insulation performance degradation curve, and generating a device-level degradation feature vector; S405, in the data holding node of the longitudinal federal learning architecture, based on the thermal spot spatial feature vector, the electrical time sequence feature vector, and the device-level degradation feature vector, performing splicing and fusion, and using an attention mechanism to allocate different weights for different types of feature vectors, and generating a fusion feature vector; S406, inputting the fusion feature vector into a fully connected layer of the fault diagnosis model, performing nonlinear transformation and feature mapping on the fusion feature by the fully connected layer, and finally outputting a fault feature vector.

7. The 10 kV distribution network fault detection method based on the Internet of Things according to claim 1, characterized in that, In step S5, constructing a three-level output model based on a biological heuristic swarm intelligence optimization algorithm and a deep Q network, performing end-to-end optimization on the fault feature vector, and generating a diagnosis result including fault type, location, and severity evaluation include: S501, constructing a three-level output model based on the fault feature vector, the three-level output model comprising a feature fusion layer, a decision optimization layer, and a diagnosis output layer, wherein the feature fusion layer adopts a gated recurrent unit network to fuse multi-source fault features; S502, initializing policy network parameters of a deep Q network using a bio-inspired swarm intelligence optimization algorithm, the bio-inspired swarm intelligence optimization algorithm adopting a particle swarm optimization algorithm to search for optimal strategy weights in a solution space by simulating swarm intelligence behaviors; S503, setting a reinforcement learning environment in the deep Q network, inputting the fault feature vector as a state, defining a fault diagnosis action space comprising three types of actions, namely, fault type classification, location positioning, and severity quantification, and constructing a reward function based on historical fault data; S504, optimizing the three-level output model based on the deep Q network through an end-to-end training process, comprising: performing dimension reduction and normalization processing on the fault feature vector in the feature fusion layer to generate a fusion feature matrix; calculating Q values of each action based on the fusion feature matrix in the decision optimization layer using the deep Q network, and selecting an optimal diagnosis action sequence based on an ε-greedy strategy; S505, analyzing the action sequence in the diagnosis output layer based on the optimal diagnosis action sequence to generate a diagnosis result comprising a fault type probability distribution, a location coordinate confidence interval, and a severity score vector.

8. The 10 kV distribution network fault detection method based on the Internet of Things according to claim 7, characterized in that, In step S503, the expression of the reward function is: ， ， ， ； ， ， ； wherein, represents the total reward value of the entire fault diagnosis task, represents the dynamic weight of the fault type classification task, represents the reward value of the fault type classification task, represents the dynamic weight of the position localization task, represents the reward value of the position localization task, represents the dynamic weight of the severity quantification task, represents the reward value of the severity quantification task, represents the total training round, represents the current training round, represents the number of different fault types that can occur in the fault diagnosis task, represents the natural base, represents the distance between the predicted position and the actual position, represents the maximum acceptable distance threshold, represents the error between the predicted severity score and the actual severity score, represents the maximum allowed error.

9. The method for fault detection of 10 kV distribution network based on Internet of Things according to claim 7, characterized in that, The severity score vector outputs low, medium, and high three-level quantification values based on a fuzzy logic evaluation system.

10. A 10 kV power distribution network fault detection system based on Internet of Things, characterized in that, The system comprises: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to implement the IoT-based 10-kilovolt distribution network fault detection method of any one of claims 1 to 9 when executing the executable instructions.

Images