An electric vehicle loop network sharing intelligent operation and maintenance method

Abstract

The application relates to the technical field of charging facility operation and maintenance, and discloses an electric vehicle looped network sharing intelligent operation and maintenance method, which comprises the following steps: a cloud server acquires monitoring data and extracts a multi-dimensional fault feature vector representing the space-time operation state of a looped network and a charging module; a detection model fusing a graph neural network and a bidirectional long short-term memory network is used to deeply fuse space and time features to output a fault probability; a fault type is determined based on a probability decision logic and a multi-modal cross-validation mechanism, and a similarity matching algorithm is used to lock a specific fault source; and a hierarchical self-healing control instruction is sent to a power distribution control module and a rectification module based on a fault level, and a derating operation, isolation bypass or full stop protection strategy is respectively executed. Through the multi-dimensional feature fusion and hierarchical response mechanism, the application effectively solves the problem of difficult DC looped network fault positioning, and guarantees system service continuity.

Description

Technical Field

[0001] This invention relates to the field of charging facility operation and maintenance technology, specifically to an intelligent operation and maintenance method for electric vehicle ring network sharing. Background Technology

[0002] Power-sharing charging pile technology based on DC bus ring network architecture is widely used due to its flexible power dispatch, but its tight electrical topology poses challenges to fault detection and maintenance. First, because the electrical characteristics of each node in the ring network are highly coupled, fault signals superimpose in time and space. Existing technologies mainly rely on single-dimensional threshold monitoring, which is difficult to effectively decouple the complex time and space operating states. As a result, the system cannot accurately distinguish between normal load fluctuations and real faults, and the identification accuracy is insufficient.

[0003] Meanwhile, existing positioning mechanisms lack cross-verification of multi-source data, making them highly susceptible to false alarms due to single sensor drift or environmental noise interference. Furthermore, they struggle to accurately pinpoint specific faulty components among numerous parallel-operating devices, hindering troubleshooting by maintenance personnel. In addition, traditional protection strategies generally employ indiscriminate shutdown and power-off modes, lacking tiered response mechanisms based on fault severity. This often results in the entire charging station experiencing service interruptions due to minor local anomalies, impacting facility uptime and service continuity. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a shared intelligent operation and maintenance method for electric vehicle ring networks. This method solves the problems in existing ring DC networks, such as low identification accuracy due to the high spatiotemporal coupling of fault characteristics, susceptibility to interference in single data source positioning and difficulty in accurately locating faulty components, and the lack of a hierarchical self-healing mechanism leading to system outages and poor service continuity caused by local faults.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a method for intelligent operation and maintenance of shared electric vehicle ring networks, comprising the following steps:

[0007] The cloud server acquires multi-source raw observation vectors collected by the monitoring sensor network, performs preprocessing on the multi-source raw observation vectors and extracts multi-dimensional fault feature vectors, which characterize the spatiotemporal operating status of the ring DC bus and the charging module.

[0008] The cloud server inputs the multidimensional fault feature vector into the fault detection model, and uses the fault detection model to perform inference based on the multidimensional fault feature vector, which fuses the spatial features of the ring network topology and the temporal features of fault evolution, to obtain a probability distribution vector for a predefined fault type.

[0009] The cloud server executes probability determination logic based on the probability distribution vector, determines the fault type by combining a multimodal cross-validation mechanism, and uses a similarity matching algorithm to pinpoint the specific fault source.

[0010] The cloud server assesses the fault level based on the fault type and the fault source, sends graded self-healing control commands to the power distribution control module and the rectifier module for different fault levels, and generates maintenance work orders.

[0011] This invention provides a method for intelligent operation and maintenance of shared electric vehicle ring networks. It has the following beneficial effects:

[0012] 1. This invention extracts multidimensional fault feature vectors covering the time domain, frequency domain, phasor, and residual, and uses a fusion architecture of graph neural network and bidirectional long short-term memory network for inference. It can simultaneously capture the spatial propagation characteristics of ring network topology and the temporal evolution law of fault signals, solving the technical problem of highly coupled and difficult-to-distinguish fault features in DC ring networks, and realizing accurate identification of complex spatiotemporal fault modes.

[0013] 2. This invention uses a dual-threshold probability judgment logic and a multimodal cross-validation mechanism to perform secondary verification of suspected faults by integrating logical instructions, physical parameters and transient traveling wave signals. Furthermore, it uses a similarity matching algorithm based on Euclidean distance to pinpoint the specific faulty device, eliminating random interference from a single data source and ensuring that the specific location and component of the fault can be accurately located under complex working conditions.

[0014] 3. This invention constructs a fault level assessment matrix and implements a graded self-healing control strategy of derated operation, isolation bypass, and full shutdown protection for minor, moderate, and severe faults, respectively. This dynamic response mechanism changes the traditional indiscriminate shutdown maintenance mode, and can maintain derated operation of equipment when dealing with non-fatal faults and ensure the smooth operation of the backbone network when isolating local faults, thereby maximizing the availability of charging services while ensuring the physical safety of the system. Attached Figure Description

[0015] Figure 1 This is an overall flowchart of the intelligent operation and maintenance method for electric vehicle ring network sharing according to the present invention;

[0016] Figure 2 This is a schematic diagram illustrating the architectural principle of the fault detection model in this embodiment of the invention;

[0017] Figure 3 This is a flowchart of the probability determination logic based on dual thresholds and multimodal cross-validation in an embodiment of the present invention;

[0018] Figure 4 This is a schematic diagram of the structure of the electric vehicle ring network shared intelligent operation and maintenance system in an embodiment of the present invention.

[0019] The components include: 1. Rectifier module; 2. Ring DC bus; 3. Charging module; 31. Power module; 4. Power distribution control module; 5. Synchronous phasor measurement unit; 6. Traveling wave sensor; 7. Discharge sensor; 8. Built-in sensor; and 9. Cloud server. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See attached document Figure 4 This invention provides an intelligent operation and maintenance method for electric vehicle ring network sharing. The method includes an intelligent operation and maintenance system for electric vehicle ring network sharing, comprising a rectifier module 1, a ring DC bus 2, multiple charging modules 3, a power distribution control module 4, a monitoring sensor network, and a cloud server 9.

[0022] Rectifier module 1 is connected to the AC power grid and configured to convert AC power into high-voltage DC power. It consists of one or more parallel rectifier cabinets forming a shared power pool. A ring DC bus 2 connects to the output of rectifier module 1, connecting all charging modules 3 within the station in a ring topology. Each charging module 3 includes multiple power modules 31, the number of which is configured as needed. Power distribution control module 4 is communicatively connected to rectifier module 1 and the multiple charging modules 3, and is configured to allocate power quotas.

[0023] The monitoring sensor network includes: a synchronization phasor measurement unit 5, located at the access section of each charging module 3, the main node of the ring DC bus 2, and the port of the power distribution control module 4, configured to collect synchronization voltage and current data; a traveling wave sensor 6, located at both ends of the cable of the ring DC bus 2, configured to capture fault transient traveling waves; a discharge sensor 7, located at the cable joint, configured to monitor insulation degradation signals; and a device-integrated sensor 8, integrated inside the charging module 3, configured to collect temperature, vibration, and voltage and current data of the power module 31.

[0024] Cloud server 9 is communicatively connected to power distribution control module 4 and configured to acquire sampling times collected by the monitoring sensor network. Multi-source raw observation vector

[0025] ;

[0026] in, and These represent the sets of synchronous voltage phasors and synchronous current phasors collected by the synchronous phasor measurement unit 5, respectively. This represents the value of the transient traveling wave signal captured by traveling wave sensor 6; This indicates the value of the insulation degradation signal monitored by discharge sensor 7; This represents the device status dataset collected by the device's built-in sensor 8, which includes temperature, vibration, module voltage, and module current. The transpose operator for vectors. Cloud server 9 is based on... The system performs fault detection calculations and sends control commands to the power distribution control module 4. The hardware implementation for the acquisition and transmission of the aforementioned data is well-known in the field and will not be described in detail here.

[0027] See attached document Figure 1 This invention provides an intelligent operation and maintenance method for electric vehicle ring network sharing. The method is executed collaboratively by a cloud server 9 and a rectifier module 1, a ring DC bus 2, a charging module 3 and a power distribution control module 4 deployed at the charging station site.

[0028] First, the acquisition and preprocessing of multi-source heterogeneous data is performed. The cloud server 9 acquires in real-time the raw observation vectors from multiple sources collected by the synchronous phasor measurement unit 5, traveling wave sensor 6, discharge sensor 7, and the device's built-in sensors 8. Given the heterogeneity of the data sources, cloud server 9 performs time-series alignment on synchronous voltage and current phasors, traveling wave signals, insulation degradation signals, and equipment status data from different locations based on a unified timestamp, and removes outliers. To reduce transmission bandwidth consumption and preserve transient features, cloud server 9 performs wavelet transform compression on the time-domain data, reconstructing key features by retaining high-amplitude wavelet coefficients. Subsequently, cloud server 9 extracts multidimensional fault feature vectors from the preprocessed data. This multidimensional fault feature vector It covers time-domain characteristics describing the statistical properties of the signal, frequency-domain characteristics describing harmonics and frequency shifts, phasor characteristics describing the phase relationship between nodes of the ring DC bus 2, and residual characteristics describing the degree of deviation of the observed values ​​from the normal reference.

[0029] Following this, the process moves into the fault simulation and identification phase based on digital twins. The cloud server 9 will extract multi-dimensional fault feature vectors. The input is fed into a pre-trained fault detection model. This fault detection model employs a fusion architecture of Graph Neural Network (GNN) and Bidirectional Long Short-Term Memory (Bi-LSTM). The GNN layer aggregates spatial features using the adjacency matrix of the ring network topology to identify the spatial propagation pattern of faults between the ring DC bus 2 and each charging module 3; the Bi-LSTM layer embeds time-series data to capture the temporal evolution patterns before and after the fault occurs. The model outputs probability distribution vectors for three predefined fault types (fault in charging module 3 or power distribution control module 4, link fault in ring DC bus 2, and power distribution logic fault). .

[0030] Next, a multi-level probability determination and precise fault location phase is performed. The cloud server 9 extracts the probability distribution vector. The highest failure type probability value and compare it with the first preset probability threshold. and the second preset probability threshold A comparison is made. In this embodiment, to balance the accuracy and recall of fault detection, a first preset probability threshold is used. The value range is set to 0.85 to 0.95 to ensure that qualitative analysis is only performed when the feature is extremely obvious; the second preset probability threshold The value range is set to 0.50 to 0.60 to filter out low-confidence results that are random guesses. Then, the following branch logic is executed:

[0031] like The cloud server 9 directly determines the type of fault that has the highest probability of occurring in the system.

[0032] like The cloud server 9 determines that the current data features are not obvious and returns to the data acquisition stage to collect data again.

[0033] like If this occurs, cloud server 9 will trigger a cross-validation mechanism for this type of fault. The cross-validation mechanism includes: logical conflict verification based on power allocation logs, device data verification based on multi-parameter mutations of the device's built-in sensors 8, and link verification based on transient traveling wave signals from the traveling wave sensor 6.

[0034] When the fault type is confirmed to be a fault of charging module 3, the cloud server 9 further calculates the Euclidean distance between the real-time monitoring data of each charging module 3 and the data of the fault sample center, and locks the charging module 3 with the smallest Euclidean distance as the specific fault source.

[0035] Finally, the hierarchical operation and maintenance strategy and closed-loop response phase are implemented. Based on the determined fault type, number of fault sources, and impact on ring network power distribution, cloud server 9 classifies faults into three levels: minor, moderate, and severe. For each level, cloud server 9 issues corresponding automatic control commands to the power distribution control module 4, including disconnecting individual charging module 3 connections, isolating local links of the ring DC bus 2, or disconnecting the rectifier module 1 from the main power supply connection of the ring DC bus 2. Simultaneously, cloud server 9 queries a pre-set associated database based on the fault type to obtain the corresponding field verification solution and pushes the solution to the user terminal to assist maintenance personnel in on-site troubleshooting.

[0036] In this embodiment, cloud server 9 performs deep cleaning, compression, and feature engineering on multi-source heterogeneous data from the monitoring sensor network to construct high-quality input vectors for model inference.

[0037] The cloud server 9 first receives the multi-source raw observation vectors uploaded by the synchronous phasor measurement unit 5, the traveling wave sensor 6, the discharge sensor 7, and the built-in sensor 8. Due to differences in data sampling frequencies and transmission delays among the various sensors, cloud server 9 uses the high-precision GPS timing timestamp of synchronization phasor measurement unit 5 as a benchmark to perform time-domain alignment processing on all data streams. During the alignment process, cloud server 9 employs a statistical filtering algorithm to identify and remove outliers caused by electromagnetic interference. Subsequently, to address the bandwidth consumption issue caused by massive high-frequency transient data transmission while retaining transient singularities crucial for fault identification, cloud server 9 uses discrete wavelet transform to compress the time-domain waveform data. Cloud server 9 utilizes the mother wavelet function... For input time domain signal Perform multi-scale decomposition and calculate wavelet coefficients. The calculation formula is as follows:

[0038] ;

[0039] in, This represents the scaling factor, used to control frequency resolution. This represents the translation factor, used to control time positioning; It is the complex conjugate of the mother wavelet function; This represents the energy normalization coefficient, used to ensure that the energy is normalized at different scale factors. Under these conditions, the energy of the transformed signal remains consistent with that of the original signal; This represents the input time-domain signal to be processed, i.e., the signal acquired by the sensor at different sampling times. The original voltage or current data that has changed; The integral symbol represents the integral operation over the entire real number line time domain (from negative infinity to positive infinity); This represents the time differential variable, i.e., the integral variable in integral operations.

[0040] The cloud server 9 sets a threshold to filter and retain wavelet coefficients with high amplitude. The threshold is configured to be 5% to 10% of the maximum modulus of the wavelet coefficients within the current time window, and low-amplitude background noise coefficients are filtered out, thereby reconstructing a compressed data sequence that retains key transient features. The specific algorithm implementation of discrete wavelet transform and the selection of the mother wavelet function can be learned by those skilled in the art using existing mathematical tool libraries. In fact, it is a well-known technology in the field and will not be described in detail here.

[0041] After completing data preprocessing, cloud server 9 extracts multidimensional fault feature vectors from the reconstructed data sequence. This vector contains feature components in four dimensions: time-domain, frequency-domain, and phasor residuals.

[0042] In the time-domain feature dimension, cloud server 9 calculates the slope of abrupt changes in the peak mean variance of the voltage and current series. Among these, the slope of the abrupt changes... The formula used to quantify the drastic changes in a signal over a very short period of time is as follows:

[0043] ;

[0044] in, and Sampling time and the previous sampling time The signal value; This represents the sampling time interval.

[0045] In the frequency domain, the cloud server performs Fast Fourier Transform on nine pairs of time-domain signals to extract the amplitudes of key subharmonics. And the offset of the main frequency signal relative to the rated frequency. This is to identify spectral anomalies caused by faults in the internal switching devices of power module 31. In terms of phasor characteristics, cloud server 9 uses data from synchronous phasor measurement unit 5 to calculate different nodes in the ring DC bus 2. With nodes Voltage or current phasor difference between The calculation formula is:

[0046] ;

[0047] in, and Representing nodes respectively and nodes At sampling time The phase angle.

[0048] phasor difference It can reflect the impedance changes and power flow distribution of ring network links. In the residual characteristic dimension, cloud server 9 runs a baseline prediction model trained on normal historical data. Cloud server 9 calculates the current measured values. The predicted value output by the baseline prediction model residual sequences between :

[0049] ;

[0050] Cloud server 9 further statistically analyzes the residual sequence The mean, peak value, and duration of these parameters are used as key indicators to determine whether early minor faults exist in the system. The cloud server 9 concatenates the aforementioned time-domain, frequency-domain, phasor, and residual features to form a multi-dimensional fault feature vector that is ultimately input into the fault detection model. .

[0051] See attached document Figure 2 In this embodiment, the cloud server 9 constructs a deep learning fault detection model based on digital twin technology. This model aims to integrate the spatial characteristics of the ring network topology with the temporal characteristics of fault evolution to achieve accurate classification of complex fault types.

[0052] This fault detection model employs a hybrid architecture combining a Graph Neural Network (GNN) and a Bidirectional Long Short-Term Memory (Bi-LSTM) network. Specifically, the cascading architecture refers to the cloud server 9 using the GNN as a pre-processor feature extractor to independently extract the spatial topological features of the ring network for each time step within the sampling time window. The extracted spatial features from each time step are then aggregated to form a time series, which is sequentially input into the Bi-LSTM for temporal inference. The cloud server 9 first constructs graph-structured data based on the physical connectivity of the electric vehicle ring network shared charging system. Among them, the node set Corresponding to each charging module 3, power distribution control module 4, and key sensor node, the edge set This corresponds to the physical link of the ring DC bus 2 that connects the aforementioned nodes.

[0053] The cloud server 9 will extract multi-dimensional fault feature vectors The features are mapped to node characteristics in a graph structure and input into a Graph Neural Network (GNN) layer. The GNN layer uses graph convolution operations to aggregate information from neighboring nodes to capture the propagation characteristics of faults in the ring network topology. The GNN layer... Network layer node feature output matrix The formula for calculating the update rule is as follows:

[0054] ;

[0055] in, Indicates the first The node feature output matrix of the network layer contains the aggregated spatial structure information; This represents a nonlinear activation function (such as the ReLU function) used to introduce nonlinear feature transformations; The adjacency matrix with added self-loops is defined as follows: ,in This is the original adjacency matrix. The identity matrix represents the fact that nodes consider both their own features and those of their neighbors when aggregating information. Representation matrix The degree matrix is ​​used to normalize the features and prevent numerical explosion. This represents a symmetric normalized adjacency matrix, which is used as a whole to balance the influence of nodes of different degrees during graph convolution. Indicates the first The node feature input matrix of the network layer; Indicates the first The learnable weight matrix of the network layer is used to perform linear transformations on the feature dimensions.

[0056] The spatial feature sequences extracted by the GNN layer are then fed into a Bidirectional Long Short-Term Memory (Bi-LSTM) layer. The Bi-LSTM layer contains two independent LSTM chains, one forward and one backward, processing the forward and reverse time series respectively to capture long-range temporal dependencies before and after the fault occurs. The Bi-LSTM layer samples at specific times... Fusion hidden state vector The calculation formula is as follows:

[0057] ;

[0058] in, This indicates that the Bi-LSTM layer is at the sampling time. The fused hidden state vector contains complete temporal context information; This indicates the forward LSTM chain at the sampling time. The hidden state captures information from the past to the present; This indicates the backward LSTM chain at the sampling time. The hidden state captures information from the future (relative to the current time step) to the present; This represents a vector concatenation operation, which merges the forward and backward states into a higher-dimensional vector.

[0059] The model's output layer is a fully connected layer connected to a Softmax classifier, used to fuse the hidden state vectors. Mapped to the probability distribution vector of fault types .

[0060] To train the above model, cloud server 9 pre-constructed a training dataset containing three major categories of fault samples. These three categories of faults are defined as follows:

[0061] Equipment-level faults include capacitor aging, heat dissipation failure, and power module open circuits occurring inside the charging module 3 or power distribution control module 4.

[0062] Link-level faults include cable breakage, insulation degradation, and loose connections occurring on the ring DC bus 2.

[0063] Logic-level faults include algorithm parameter drift, logic conflicts, and command transmission delays occurring in the power distribution control module 4.

[0064] For each specific fault scenario, cloud server 9 collects data including the time of fault occurrence. The first 100ms and the time of fault occurrence The last 500ms data segment is used as a set of standard fault samples. Cloud server 9 collects at least 1000 fault samples for each fault type to form a balanced training dataset. During model training, the learnable weight matrix is ​​optimized using the cross-entropy loss function and backpropagation algorithm. And LSTM parameters. The specific layer settings, learning rate hyperparameter adjustments, and backpropagation algorithm implementations for graph neural networks and LSTM networks can be configured by those skilled in the art based on actual computing resources; these are well-known techniques in the field and will not be elaborated upon here.

[0065] See attached document Figure 3 In this embodiment, the cloud server 9 outputs the probability distribution vector of the fault detection model. Multi-level threshold adjudication is performed, and a multi-modal cross-validation mechanism is combined to achieve final confirmation of the fault type and physical location locking.

[0066] Cloud server 9 first parses the probability distribution vector output by the fault detection model. ,in, These correspond to the confidence levels of "charging module 3 or power distribution control module 4 failure", "ring network link failure", and "power distribution logic failure", respectively. The probability distribution vector is extracted from cloud server 9. The maximum value in the range is used as the probability value of the highest failure type. And identify the fault type index corresponding to the maximum value as the undetermined fault type. .

[0067] Cloud server 9 Compared with the preset first preset probability threshold (High confidence threshold) and second preset probability threshold (For low confidence threshold) numerical comparisons, execute the following branch logic:

[0068] like This indicates that the model has extremely high confidence in identifying fault types, and the cloud server 9 directly determines the current fault occurrence in the system. The type of fault is identified, and the process proceeds directly to the fault location stage.

[0069] like This indicates that the current feature data is insufficient to support a clear conclusion about the fault, due to transient interference or data loss. Cloud server 9 determines this detection is invalid and sends a command to control the monitoring sensor network to re-acquire multi-source raw observation vectors for the next time window. .

[0070] like This indicates the presence of suspected fault characteristics but not meeting the certainty criteria; cloud server 9 triggers a response for the pending fault type. The cross-validation mechanism.

[0071] when When a "power distribution logic fault" occurs, cloud server 9 performs logic verification: it retrieves the control log of power distribution control module 4 to check for the existence of a time window. Inside, a certain charging module 3 issued a non-zero power request. However, the actual power allocation command issued by the power allocation control module 4... Furthermore, no hardware alarms were recorded during this period. If the above logical conflict conditions are met, a power distribution logic fault is confirmed to have occurred.

[0072] when When the error message is "Fault in charging module 3 or power distribution control module 4", the cloud server 9 performs data verification: retrieves temperature data from the device's built-in sensor 8. and vibration data The system determines whether the rate of change exceeds the physical safety threshold. The physical safety threshold is set according to the equipment's manufacturer's specifications; for example, the temperature change rate threshold is set to rise by 5°C to 10°C per minute, and the vibration acceleration threshold is set to 2g to 5g (g is gravitational acceleration). If there is an abnormal sudden change in the physical parameters, an equipment-level fault is confirmed.

[0073] when When the fault is identified as a "ring network link failure", cloud server 9 performs link verification: it checks the transient traveling wave signal collected by traveling wave sensor 6. Does the amplitude exceed the background noise threshold? Background noise threshold It is not a fixed value, but rather the average amplitude of traveling wave sensor data during periods of no load or light load, calculated by the cloud server 9 statistical system, and a dynamic threshold is set at 1.5 to 2 times this average amplitude. If a distinct traveling wave front is detected, a ring network link failure is confirmed.

[0074] After confirming the fault type as "charging module 3 fault," cloud server 9 needs to further pinpoint the specific faulty unit from multiple charging modules 3. Cloud server 9 executes a similarity matching algorithm based on Euclidean distance. Cloud server 9 extracts the real-time normalized operating parameters of all charging modules 3. Specifically, the real-time normalized operating parameters are calculated based on data collected by the device's built-in sensors 8, mainly including the DC output voltage ripple coefficient (used to characterize the health status of the filter capacitor), the unit power temperature rise rate (used to characterize the heat dissipation system status), and the module energy conversion efficiency. Cloud server 9 performs maximum-minimum normalization processing on the above physical quantities to eliminate the differences in numerical dimensions caused by the different power allocation quotas currently undertaken by each charging module 3, thereby constructing a state vector that only reflects the characteristics within the device. Subsequently, cloud server 9 retrieves the standard sample center vector of this type of fault from the fault database. Cloud server 9 calculates the value of each charging module... The Euclidean distance between the real-time data vector and the fault sample center vector The calculation formula is as follows:

[0075] ;

[0076] in, Indicates the first The Euclidean distance between the real-time status and the standard fault status of each charging module 3; the smaller the value, the higher the matching degree. This represents the total number of feature dimensions involved in the calculation (e.g., the total number of features such as temperature, voltage, and current ripple coefficient). The index variable represents the feature dimension, traversing from 1 to... ; Indicates the first The charging module 3 is in the first Real-time normalized values ​​across each feature dimension; This indicates that the center of the standard fault sample cluster in the database is at the [number]th [location]. Statistical average across each feature dimension; This indicates that the summation operation is performed on the squared differences of all feature dimensions.

[0077] about For the acquisition and definition of data, cloud server 9 employs a density-based clustering algorithm. Specifically, cloud server 9 pre-extracts all historical normalized sample vectors of the same fault type (e.g., "charging module capacitor aging") from the fault database. To remove noise interference from historical data, cloud server 9 performs K-Means clustering on these samples to identify the core cluster with the highest sample density. Subsequently, cloud server 9 calculates the density of all samples in this core cluster at the [missing information - likely a specific time point]. The arithmetic mean of the nth feature dimension is used as the nth feature center vector of the standard fault sample. Dimensional components This cluster center-based method can effectively eliminate the bias of outliers on the baseline vector, ensuring the robustness of fault location.

[0078] The cloud server 9 iterates through all charging modules 3 within the site to calculate the Euclidean distance. And select the value with the smallest Euclidean distance. Charging module 3 is identified as the final faulty charging module. The specific programming implementation of the above logical judgments and numerical comparisons can be achieved by those skilled in the art using existing computer logic control statements and mathematical operation libraries; these are well-known technologies in the field and will not be elaborated upon here.

[0079] In this embodiment, the cloud server 9 executes a graded handling strategy based on the fault type, fault source location, and the current real-time operating status of the system determined in the aforementioned steps, and generates a closed-loop operation and maintenance work order.

[0080] First, cloud server 9 establishes a fault level assessment matrix. Cloud server 9 obtains the impact factors of the fault source on the voltage stability of the ring DC bus 2 and the number of affected charging modules 3. Based on this information, cloud server 9 classifies the fault into three levels:

[0081] The first level is a minor fault. This level corresponds to a fault in an individual power module 31 within a single charging module 3, or a data outlier in a single sensor in the monitoring sensor network, and the bus voltage fluctuation of the ring DC bus 2 does not exceed a preset percentage of the rated value (e.g., 5%).

[0082] The second level is a moderate fault. This level corresponds to the complete failure of a single charging module 3, a logical conflict in the power distribution control module 4 that does not affect electrical safety, or a decrease in the insulation performance of a local branch of the ring DC bus 2 that does not cause a short circuit. At this time, the system still has the ability to maintain partial load operation.

[0083] The third level is a severe fault. This level corresponds to the failure of rectifier module 1, a short circuit or open circuit in the main trunk of the ring DC bus 2, or the detection of thermal runaway precursors that could lead to a fire. At this point, the system no longer meets the conditions for safe operation.

[0084] For different fault levels, cloud server 9 generates differentiated automatic control commands and sends them to power distribution control module 4 and rectifier module 1 to execute the following self-healing response:

[0085] For a Level 1 fault, cloud server 9 executes a derating strategy. Cloud server 9 sends a power limiting command to power distribution control module 4, adjusting the output power limit of the faulty charging module 3 to the sum of the total rated power of the remaining healthy power modules 31. Simultaneously, cloud server 9 marks the faulty power module 31 as unavailable, prohibiting subsequent power allocation tasks to it.

[0086] For Level 2 faults, cloud server 9 executes an isolation bypass strategy. If the fault source is a single charging module 3, cloud server 9 sends a disconnect command to power distribution control module 4, controlling the DC contactor connecting charging module 3 to the ring DC bus 2 to disconnect it, physically isolating it from the ring network. If the fault source is a power distribution logic conflict, cloud server 9 sends a reset command to power distribution control module 4, forcing the default parameter configuration of the reload control algorithm.

[0087] In response to a Level 3 fault, cloud server 9 executes a full shutdown protection strategy. Cloud server 9 sends an emergency shutdown command to rectifier module 1, cutting off the input power to the AC grid. Simultaneously, cloud server 9 sends trip commands to the sectionalizing circuit breakers on the ring DC bus 2, dividing the ring network into multiple de-energized sections to prevent the spread of fault current.

[0088] While executing the automatic control strategy, the cloud server 9 initiates a human-machine collaborative operation and maintenance process. Based on a pre-configured maintenance knowledge base indexed by the defined fault types, the cloud server 9 generates a digital work order containing fault codes, fault location coordinates (down to the station number or cable segment), a recommended list of spare parts (e.g., power module model, fuse specifications), and troubleshooting steps. The cloud server 9 then pushes this digital work order to the handheld terminal device of the operation and maintenance personnel via a wireless network.

[0089] After maintenance personnel arrive on-site and complete the repair, they upload a repair confirmation signal to cloud server 9 via a handheld terminal device. Upon receiving the signal, cloud server 9 controls the monitoring sensor network to perform a reconfirmation scan of the repaired area. Only when the reconfirmation scan results show that all indicators have returned to normal ranges will cloud server 9 release the fault lockout state and restore normal power distribution and operation of the system. The encoding and transmission of the above control commands and database query calls can be implemented by those skilled in the art using existing industrial communication protocols and database management systems, which are well-known technologies in the field and will not be elaborated upon here.

[0090] Meanwhile, cloud server 9 executes an adaptive update process for the fault database. Cloud server 9 marks the observation vectors at the moment of the confirmed fault occurrence, verified by on-site maintenance personnel, as high-confidence ground truth samples and records them into the corresponding fault category in the fault database. Whenever the number of new samples reaches a preset threshold (e.g., every 50 new cases), cloud server 9 triggers an update calculation, re-executing the aforementioned K-Means clustering and arithmetic mean operations to dynamically update the statistical average for that fault type. Through this closed-loop feedback mechanism, the system can continuously correct the fault baseline over time and adapt to parameter drift caused by equipment aging.

[0091] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A shared intelligent operation and maintenance method for electric vehicle ring networks, characterized in that, This includes the following steps performed by the cloud server (9): The multi-source raw observation vectors collected by the monitoring sensor network are obtained, and the multi-source raw observation vectors are preprocessed and multi-dimensional fault feature vectors are extracted. The multi-dimensional fault feature vectors represent the spatiotemporal operating status of the ring DC bus (2) and the charging module (3). The multidimensional fault feature vector is input into the fault detection model. The fault detection model uses the multidimensional fault feature vector to fuse the spatial features of the ring network topology and the temporal features of fault evolution to perform inference and obtain the probability distribution vector for the predefined fault type. Based on the probability distribution vector, a probability determination logic is executed, and a multimodal cross-validation mechanism is used to determine the fault type. A similarity matching algorithm is then used to pinpoint the specific fault source. The fault level is assessed based on the fault type and the fault source. For different fault levels, graded self-healing control instructions are sent to the power distribution control module (4) and the rectifier module (1), and maintenance work orders are generated.

2. The method for intelligent operation and maintenance of electric vehicle ring network sharing according to claim 1, characterized in that, In the step of preprocessing the multi-source original observation vectors and extracting multi-dimensional fault feature vectors, the preprocessing includes performing time alignment and wavelet transform compression. The multidimensional fault feature vector includes time-domain features, frequency-domain features, phasor features, and residual features; The time-domain features are obtained by calculating the peak value, mean, variance, and abrupt change slope of the voltage and current sequences. The frequency domain features are obtained by performing a fast Fourier transform on the time domain signal to extract the amplitude of the key subharmonics and the offset of the main frequency signal relative to the rated frequency. The data collected by the synchronous phasor measurement unit (5) in the phasor characteristic monitoring sensor network is obtained by calculating the absolute value of the difference between the voltage or current phase angles between different nodes in the ring DC bus (2). The residual feature is obtained by calculating the absolute value of the difference between the current measured value and the predicted value output by the benchmark prediction model.

3. The method for intelligent operation and maintenance of electric vehicle ring network sharing according to claim 1, characterized in that, The fault detection model adopts a fusion architecture of graph neural network and bidirectional long short-term memory network; The graph neural network aggregates the spatial features using the adjacency matrix of the ring network topology to identify the spatial propagation pattern of the fault on the ring DC bus (2). The bidirectional long short-term memory network captures the temporal features by splicing forward chain hidden states and backward chain hidden states, which is used to identify the temporal evolution pattern of faults.

4. The method for intelligent operation and maintenance of electric vehicle ring network sharing according to claim 1, characterized in that, The execution probability determination logic includes: Extract the highest fault type probability value from the probability distribution vector, and compare the highest fault type probability value with a first preset probability threshold and a second preset probability threshold. If the probability value of the highest fault type is greater than the first preset probability threshold, the fault type corresponding to the probability value of the highest fault type in the system is directly determined. If the probability value of the highest fault type is less than the second preset probability threshold, the detection is determined to be invalid and the monitoring sensor network is controlled to re-collect data.

5. The method for intelligent operation and maintenance of electric vehicle ring network sharing according to claim 4, characterized in that, If the probability value of the highest fault type is between the second preset probability threshold and the first preset probability threshold, the multimodal cross-validation mechanism is triggered. The multimodal cross-validation mechanism includes: When the pending fault type is a power allocation logic fault, verify whether there is a logical conflict where a non-zero power request and a zero power allocation command coexist without a hardware alarm. When the pending fault type is a fault of the charging module (3) or the power distribution control module (4), verify whether the rate of change of physical parameters collected by the device built-in sensor (8) in the monitoring sensor network exceeds the physical safety threshold. When the pending fault type is a ring network link fault, verify whether the amplitude of the transient traveling wave signal collected by the traveling wave sensor (6) in the monitoring sensor network exceeds the background noise threshold. When the pending fault type is an insulation fault, verify whether the intensity of the partial discharge signal collected by the discharge sensor (7) in the monitoring sensor network exceeds the preset discharge threshold.

6. The method for intelligent operation and maintenance of electric vehicle ring network sharing according to claim 1, characterized in that, The steps for using a similarity matching algorithm to pinpoint the specific fault source include: When the fault type is confirmed as a fault of the charging module (3), extract the real-time normalized operating parameters of all charging modules (3) and the center vector of the standard fault sample in the fault database; The standard fault sample center vector is obtained by performing clustering operations on historical sample vectors of the same type of fault in the fault database to identify core clusters, and calculating the arithmetic mean of all samples in the core cluster. Calculate the Euclidean distance between the real-time normalized value of each charging module (3) and the center vector of the standard fault sample, and select the charging module (3) with the smallest Euclidean distance value as the fault source.

7. The method for intelligent operation and maintenance of electric vehicle ring network sharing according to claim 1, characterized in that, The steps for assessing the fault level based on the fault type and the fault source include establishing a fault level assessment matrix: A minor fault is defined as a failure of the internal power module (31) of a single charging module (3) or a case of outlier data from a single sensor in the monitoring sensor network. The failure of a single charging module (3) as a whole, power distribution logic conflict, or local branch insulation degradation is classified as a medium fault. The failure of the rectifier module (1) and the short circuit or open circuit of the main trunk of the ring DC bus (2) are classified as serious faults.

8. The method for intelligent operation and maintenance of electric vehicle ring network sharing according to claim 7, characterized in that, For the minor fault, the graded self-healing control instruction includes a derating operation instruction, which is configured to control the power distribution control module (4) to adjust the upper limit of the output power of the charging module (3) where the fault is located to the sum of the total rated power of the remaining healthy power modules (31).

9. A shared intelligent operation and maintenance method for electric vehicle ring networks according to claim 7, characterized in that, For the aforementioned moderate fault, the graded self-healing control command includes an isolation bypass command; If the fault source is a single charging module (3), the isolation bypass command is configured to control the power distribution control module (4) to disconnect the charging module (3) from the ring DC bus (2); If the fault source is a power allocation logic conflict, the isolation bypass instruction is configured to control the default parameter configuration of the overload control algorithm of the power allocation control module (4).

10. A shared intelligent operation and maintenance method for electric vehicle ring networks according to claim 7, characterized in that, In response to the severe fault, the graded self-healing control command includes a full stop protection command, which is configured to control the rectifier module (1) to cut off the AC grid side input power and control the sectional circuit breaker of the ring DC bus (2) to trip. After the cloud server (9) receives the repair confirmation signal and the reconfirmation scan result of the monitoring sensor network shows that the indicators have returned to normal, the fault lock state is released.

Images