Network fault diagnosis model library based on cloud native and end cloud network cooperative positioning method

Abstract

The invention discloses a network fault diagnosis model library based on cloud native and an end cloud network cooperative positioning method, and relates to the technical field of network communication and intelligent diagnosis. The model library comprises a basic model construction module and a model management and iteration module. The basic model construction module constructs a basic fault diagnosis model by combining an orthogonal optimization design method with a convolutional neural network, trains the model by using historical fault data, and optimizes model parameters through a range analysis method and a variance analysis method. Carrying out differential modeling according to fault types; and the model management and iteration module packages the optimized model into a container mirror image, manages a version through a version control system, records iteration information, and supports model gray release and dynamic update based on a cloud native automatic deployment mechanism. Through an end-edge-cloud collaborative architecture, an optimized model library and a hierarchical processing mechanism, efficient response and accurate positioning of network fault diagnosis are realized, and an intelligent and automatic fault diagnosis solution is provided for a complex network scene.

Description

Technical Field

[0001] The present invention relates to the field of network communication and intelligent diagnosis technology. More specifically, the present invention relates to a cloud-native network fault diagnosis model library and an end-cloud-network collaborative positioning method. Background Art

[0002] As networks expand and their architectures become more complex, traditional fault diagnosis methods have the following flaws: reliance on manual experience leads to low efficiency and makes it difficult to cope with the dynamically changing cloud-native environment; data collection is decentralized and pre-processing capabilities are insufficient, resulting in low multi-source data fusion; there is a lack of end-edge-cloud collaboration mechanisms, resulting in high fault location delays.

[0003] Existing technologies lack a standardized approach to combining cloud-native technologies with AI. Model training and deployment remain insufficiently automated, and collaboration between devices, edge devices, and the cloud remains largely limited to data forwarding, lacking deep fault diagnosis collaboration. Therefore, an integrated solution combining model library construction and multi-device collaborative positioning is urgently needed to improve the accuracy and real-time nature of fault diagnosis. Summary of the Invention

[0004] The present invention aims to provide a cloud-native network fault diagnosis model library and an end-cloud-network collaborative positioning method to solve the problems of low diagnostic efficiency, poor model reusability, and high collaborative positioning delay in traditional methods, and to achieve rapid identification, accurate positioning, and efficient repair of faults.

[0005] To achieve the above object, the present invention provides the following technical solutions: Cloud-native network fault diagnosis model library, including: The basic model construction module uses the orthogonal optimization design method combined with the convolutional neural network to build a basic fault diagnosis model. The model is trained using historical fault data and the model parameters are optimized through range analysis and variance analysis. When building the model, the modeling method is selected according to the fault type. The decision tree algorithm is used to model abnormal equipment settings, and the CNN algorithm is used to model abnormal network status. The input features are electrical quantity features. The model management and iteration module encapsulates the optimized model as a container image, manages the model version through the version control system, and records the training data and performance indicators of each iteration. Based on the cloud-native automated deployment mechanism, it supports grayscale release and dynamic updates of the model.

[0006] As a preferred method, the end-cloud-network collaborative positioning method adopts the above-mentioned cloud-native-based network fault diagnosis model library, including the following steps: Step 1: Initial diagnosis and information exchange on the device side: The device-deployed perception module collects network status data and device operating parameters, uses a local lightweight model to make a preliminary fault diagnosis, compares the device operating parameters and network status data with preset fault thresholds, and marks the fault type. If the number of neighboring nodes is greater than or equal to a preset condition, it exchanges information with the neighboring nodes and calculates the likelihood function. Otherwise, it uses the Kalman filter algorithm to generate location information. Step 2: Narrowing the edge layer fault scope: Initial fault data is aggregated and preprocessed on the edge node receiving side. Combined with locally stored network topology information, the fault propagation model is used to narrow the fault scope. The fault suspicion of edge nodes is calculated using a fault propagation probability model. The edge nodes are ranked from high to low in terms of their suspected faults. The fault suspicion and propagation paths of a pre-set number of edge nodes / links are taken as the processing results and uploaded to the cloud. Step 3: Cloud-based in-depth analysis and fault location: The cloud calls the model in the fault diagnosis model library, performs in-depth analysis on the fault feature data uploaded by the edge node, locates the root cause of the fault through multi-model fusion reasoning, and generates results including the cause of the fault and the repair plan, which are fed back to the edge node and end-side device.

[0007] Preferably, the system architecture for implementing the above-mentioned model library and the above-mentioned collaborative positioning method includes: The device-side layer consists of terminal devices with data collection and preliminary diagnostic capabilities, monitoring network status and device parameters through built-in sensing modules; The edge layer deploys edge servers responsible for receiving end-side data, pre-processing, local fault analysis, and interacting with the end-side and cloud. The cloud layer includes cloud servers and a fault diagnosis model library. The cloud servers provide computing resources to support model training, version management, and global fault location. The model library stores optimized containerized models.

[0008] Preferably, the orthogonal optimization design method is specifically as follows: Determine the factors and levels, including convolution kernel size, learning rate, and batch size. Establish an orthogonal experimental parameter table to obtain model performance data for different parameter combinations. Use the range analysis method to calculate the performance mean and range of each factor at different levels. Use the variance analysis method to calculate the total sum of squares, factor sum of squares, and error sum of squares. Judge the significance of the factors through the mean square and significance evaluation value. Combine the range and variance analysis results to select the optimal level of each factor to form the optimal parameter combination.

[0009] Preferably, when constructing a basic fault diagnosis model, the fault type is identified, including abnormal device setting value and abnormal network status; if the fault type is abnormal device setting value, a decision tree algorithm is used for modeling; if the fault type is abnormal network status, a convolutional neural network is used for modeling; Identify electrical quantity characteristics from equipment operation data, including three-phase voltage RMS, current phase angle, current direct-axis / quadrature-axis components, and zero-sequence current; and use the electrical quantity characteristics as model input.

[0010] Preferably, the local lightweight model specifically includes a decision tree or a naive Bayesian model; the preliminary fault judgment is performed by comparing the equipment operating parameters with the preset fault thresholds corresponding to the parameters. When the value of the equipment operating parameters is greater than or equal to the preset fault thresholds corresponding to the parameters, it indicates that there is a potential fault in the corresponding parameters, and the parameters are recorded as equipment fixed value abnormalities; At the same time, by comparing the parameters in the network status data with the preset fault thresholds corresponding to the parameters, when the parameters in the network status data are greater than or equal to the preset fault thresholds corresponding to the parameters, it indicates that there is a potential fault in the corresponding parameters, and the parameters are recorded as network status abnormalities.

[0011] Preferably, the multi-model fusion reasoning is used to locate the root cause of the fault, and the specific method is as follows: Input the fault feature data uploaded by the edge node into the convolutional neural network, XGBoost and Bayesian network; the fault feature data includes electrical quantity features and topological correlation features; The convolutional neural network extracts characteristic nonlinear patterns through multi-layer convolution kernels and outputs the first fault probability; XGBoost outputs the second failure probability based on gradient boosting decision tree ensemble learning; The Bayesian network infers the probability of fault conditions through a probabilistic graphical model and outputs the third fault probability; The weight optimization algorithm that minimizes the fusion error is used to determine the weight of each model. The first to third fault probabilities are fused through the weighted sum formula to obtain the fault judgment probability. .

[0012] Preferably, the likelihood function is used to describe the probability of occurrence of measurement values ​​under a given fault state, the measurement values ​​including the distance measurement values ​​between devices and the zero-sequence current values; the likelihood function adopts a Gaussian distribution model, whose mean is the theoretical reference value of the measurement value and the variance is the variance of the measurement noise.

[0013] The technical effects and advantages of the present invention are as follows: 1. This invention utilizes a three-tiered end-edge-cloud collaboration mechanism. On the end, a local lightweight model is used to initially diagnose and label fault types, reducing invalid data uploads. The edge layer integrates network topology information to narrow the fault scope, reducing cloud processing pressure. The cloud uses a model library for in-depth analysis to achieve precise location, significantly reducing manual intervention. This cloud-native architecture supports containerized model deployment, version management, and dynamic updates. By implementing grayscale releases to avoid upgrade delays, this significantly improves fault diagnosis response speed and automation.

[0014] 2. The model library of the present invention implements differentiated modeling for different fault types, using decision trees for abnormal equipment settings and CNNs for abnormal network status, and inputting electrical quantity characteristics to improve targeting. Model parameters are optimized through orthogonal optimization design and range and variance analysis to improve basic model performance. Multi-model fusion reasoning is used in the cloud, combined with the fusion results of the weight optimization algorithm, to effectively improve fault location accuracy and enhance the adaptability to diverse faults in complex network environments.

[0015] 3. This invention reduces remote data transmission volume through localized preliminary diagnosis on the end side, preprocesses data at the edge layer and narrows the scope of faults, thereby alleviating pressure on cloud computing and network bandwidth. Its layered architecture brings data processing close to the source, improving the real-time nature of fault response. Furthermore, its model library supports dynamic iteration and version management, enabling rapid adaptation to changes in network topology and fault types, and adapting to dynamic and complex network environments.

[0016] In summary, through the end-edge-cloud collaborative architecture, optimized model library and layered processing mechanism, efficient response, precise positioning and strong environmental adaptability of network fault diagnosis are achieved, effectively solving the problems of traditional methods such as reliance on manual labor, low efficiency and poor versatility, and providing intelligent and automated fault diagnosis solutions for complex network scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a principle block diagram of Example 1 of the cloud-native network fault diagnosis model library proposed by the present invention; Figure 2 This is a flowchart of Example 2 of the end-cloud-network collaborative positioning method proposed by the present invention; Figure 3 This is a principle block diagram of Example 3 of the system framework of the model library and collaborative positioning method proposed in the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] Example 1

[0020] like Figure 1 As shown in the figure, the cloud-native network fault diagnosis model library includes: The basic model construction module uses the orthogonal optimization design method combined with the convolutional neural network to build a basic fault diagnosis model. The model is trained using historical fault data and the model parameters are optimized through range analysis and variance analysis. When building the model, the modeling method is selected according to the fault type. The decision tree algorithm is used to model abnormal equipment settings, and the CNN algorithm is used to model abnormal network status. The input features are electrical quantity features. The model management and iteration module encapsulates the optimized model as a container image, manages the model version through the version control system, and records the training data and performance indicators of each iteration. Based on the cloud-native automated deployment mechanism, it supports grayscale release and dynamic updates of the model.

[0021] In this application, the orthogonal optimization design method is specifically: Determine the factors and levels, including convolution kernel size, learning rate, and batch size. Establish an orthogonal experimental parameter table, obtain model performance data under different parameter combinations through experiments, and use the range analysis method to calculate the range of each factor: The accuracy of identifying the kth experiment under any parameter combination is recorded as , calculate the performance mean of the i-th factor at the j-th level, the formula is expressed as ; Where m represents the number of experiments at this level, i represents the index of the factor, and j represents the level index of the factor; Then obtain the performance mean of the i-th factor at the 1st, 2nd and 3rd levels and record them as , calculate the range , the formula is ;in, It represents the minimum and maximum values ​​of the performance mean of the i-th factor at the 1st, 2nd and 3rd levels; The significance of factors on model performance was tested by analysis of variance: The total number of identification experiments is recorded as N, and the average accuracy of each experiment is recorded as , calculate the total sum of squares of deviations , the formula is , the sum of squares of factor deviations , the formula is , the sum of squared error deviations , the formula is ; Then calculate the mean square 、 ,in =2 is the degree of freedom of the factor, is the error degree of freedom; finally calculate the significance evaluation value of factor i , the formula is ; The significant evaluation value is greater than its preset critical value ( ), indicating that the factor has a significant impact; Combining the results of range and variance analysis, the optimal level of each factor is selected to form the optimal parameter combination.

[0022] Specifically, through orthogonal optimization design, by clarifying key factors and levels such as convolution kernel size, combining range analysis to quantify the degree of factor influence and variance analysis to test the significance of the influence, the optimal parameter combination is accurately screened, the model performance stability and generalization ability are improved, laying the foundation for adapting efficient models to different fault types, and enhancing the accuracy and reliability of fault diagnosis.

[0023] In this application, when constructing a basic fault diagnosis model, the fault type is identified, including abnormal device settings and abnormal network status. If the fault type is abnormal device settings, a decision tree algorithm is used for modeling; if the fault type is abnormal network status, a convolutional neural network is used for modeling. Identify electrical quantity characteristics from equipment operation data, including three-phase voltage RMS, current phase angle, current direct-axis / quadrature-axis components, and zero-sequence current; and use the electrical quantity characteristics as model input.

[0024] Specifically, this application differentiates the modeling of two types of faults: equipment constant value abnormalities (clear cause and effect) and network status abnormalities (complex nonlinearity). Combined with the input of electrical quantity characteristics such as the effective value of three-phase voltage, it not only realizes the rapid rule matching of constant value abnormalities through decision trees, but also uses CNN to capture the nonlinear characteristics of network status abnormalities, thereby improving the model's adaptability and diagnostic accuracy for different fault types and reducing misjudgments and missed judgments.

[0025] Example 2

[0026] Combine Figure 2 This embodiment proposes a device-cloud-network collaborative positioning method, which is implemented based on the cloud-native network fault diagnosis model library of Example 1. This embodiment is a preferred application mode of Example 1, and its implementation does not affect the independent implementation of Example 1.

[0027] Specifically, the end-cloud-network collaborative positioning method of this embodiment includes the following steps: Step 1: Initial diagnosis and information exchange on the device side: A perception module is deployed on the device side to collect network status data and device operating parameters. For devices on the device side, such as smart meters and industrial sensors, the perception module consists of hardware and software. The hardware includes voltage / current sensors and 5G communication modules, and the software includes a data collection program. It uses a local lightweight model to perform preliminary fault diagnosis, compares device operating parameters and network status data with preset fault thresholds, and marks the fault type. When the number of neighboring nodes is greater than or equal to a preset condition, it exchanges information with the neighboring nodes and calculates the likelihood function. Otherwise, it uses the Kalman filter algorithm to generate location information. The specific implementation of using the Kalman filter algorithm to generate location information refers to existing mature methods. For example, by constructing state transfer equations and observation equations, using a prediction-update iterative process, fusing the device motion state with neighboring node ranging data, and outputting a high-precision location estimate. This algorithm has been widely used in wireless sensor network positioning scenarios. This application uses it to achieve real-time solution of the terminal-side device location, providing basic data for fault spatial correlation analysis. Step 2: Narrowing the edge layer fault scope: The edge node receives preliminary fault data and performs preprocessing. Combined with the locally stored network topology information, the fault propagation model is used to narrow the fault scope. The fault suspicion of the edge node is calculated using the fault propagation probability model. , the formula is Where, is the set of neighbor nodes of node v, where v represents the index of the edge node. is the link fault propagation weight from node u to v, which can be trained based on factors such as link impedance and historical fault frequency; The fault suspicion of all edge nodes is sorted from high to low. The fault suspicion and propagation paths of a predetermined number of edge nodes / links are taken as the processing results and uploaded to the cloud. Preprocessing includes interpolation to fill in null values, calculating per-unit values, and normalization. The network topology information is abstracted into a weighted graph G = (V, E, W), where V is the set of device nodes, E is the set of links, and W is the link fault propagation weight. Step 3: Cloud-based in-depth analysis and fault location: The cloud calls the model in the fault diagnosis model library, performs in-depth analysis on the fault feature data uploaded by the edge node, locates the root cause of the fault through multi-model fusion reasoning, and generates results including the cause of the fault and the repair plan, which are fed back to the edge node and end-side device.

[0028] Specifically, the terminal side improves positioning efficiency through preliminary diagnosis and interaction with neighbors using lightweight models. The edge layer reduces the pressure on the cloud by shrinking the scope through data preprocessing and fault propagation models. The cloud-side multi-model fusion realizes precise positioning and feedback solutions. The three-level collaborative mechanism greatly improves the real-time and accuracy of fault diagnosis, effectively solving the problems of lack of coordination and high positioning delay in traditional methods.

[0029] In this application, the local lightweight model specifically includes a decision tree or a naive Bayesian model; the preliminary fault judgment is performed by comparing the device operating parameters with the preset fault thresholds corresponding to the parameters. When the value of the device operating parameter is greater than or equal to the preset fault threshold corresponding to the parameter, it indicates that the corresponding parameter has a potential fault, and the parameter is recorded as a device fixed value abnormality; At the same time, by comparing the parameters in the network status data with the preset fault thresholds corresponding to the parameters, when the parameters in the network status data are greater than or equal to the preset fault thresholds corresponding to the parameters, it indicates that there is a potential fault in the corresponding parameters, and the parameters are recorded as network status abnormalities.

[0030] Specifically, lightweight models such as decision trees or naive Bayes are used to adapt to the low-computing power environment on the terminal side. By comparing parameters with preset thresholds, abnormal device values ​​and network status are quickly marked. This not only achieves preliminary screening and classification of faults and reduces invalid data uploads, but also provides accurate fault type labels for subsequent in-depth diagnosis of the edge layer and cloud, thereby improving the efficiency and pertinence of collaborative positioning.

[0031] In this application, multi-model fusion reasoning is used to locate the root cause of the fault. The specific method is as follows: The fault feature data uploaded by the edge node is input into the convolutional neural network, XGBoost, and Bayesian network. The fault feature data includes electrical quantity features and topological association features. The topological association features are features related to fault propagation extracted based on the network topology structure (abstracted as a weighted graph G = (V, E, W)). Specifically, they include node connectivity, link weight attributes, fault propagation path features, and topological hierarchical relationships. Node connectivity represents the number of neighbor nodes of the device node v, reflecting the closeness of the node's connection in the network. The link weight attribute represents the fault propagation weight W(u, v) of the link between nodes, which is obtained through normalization of quantitative indicators such as link impedance, historical fault frequency, and bandwidth utilization. The fault propagation path features represent the shortest path length from the suspected source node to the target node and the number of key nodes passed through, depicting the propagation trajectory of the fault in the topology. The topological hierarchical relationship represents the level of the node in the network layered architecture and the size of the impact range of the associated fault. The convolutional neural network extracts characteristic nonlinear patterns through multi-layer convolution kernels and outputs the first fault probability P1; XGBoost is based on gradient boosting decision tree ensemble learning and outputs the second failure probability P2; The Bayesian network infers the probability of fault conditions through the probabilistic graphical model and outputs the third fault probability P3; The weight optimization algorithm that minimizes the fusion error is used to determine the weights of each model, which are recorded as p1, p2, and p3 respectively. The first to third fault probabilities are fused through the weighted sum formula. The formula is: , get the fault judgment result ; The objective function of weight optimization is to minimize the mean square error (MSE) between the fusion result and the true fault label. The formula is: ; Where M is the number of samples in the validation set, is the true fault label of the h-th sample, h represents the index of the sample in the validation set, It is represented as the fault judgment result output of the hth sample after multi-model fusion; Through the iterative optimization algorithm, we search for p1, p2, and p3 that satisfy the constraints to minimize the mean square error. The weight corresponding to the minimized mean square error is used as the final effective weight for multi-model fusion reasoning.

[0032] Specifically, convolutional neural networks, XGBoost, and Bayesian networks are used to capture the nonlinear patterns, integrated learning rules, and conditional probability relationships of fault characteristics, respectively. The output is then fused with a weighted optimization algorithm that minimizes the mean square error. This not only complements the advantages of each model, but also improves the stability and accuracy of fault judgment through weighted summation and iterative optimization, providing a more reliable quantitative basis for fault root cause location.

[0033] In this application, the likelihood function is used to describe the probability of occurrence of measurement values ​​under a given fault state. The measurement values ​​include the distance measurement value d between devices and the zero-sequence current value. ; The likelihood function adopts the Gaussian distribution model, and its mathematical expression is: ;in is the fault status parameter, is the theoretical reference value of the distance measurement between devices. is the variance of the distance measurement noise, is the theoretical reference value of the zero-sequence current value, is the variance of the zero-sequence current measurement noise.

[0034] Specifically, the likelihood function is based on a Gaussian distribution model. It combines the measured values ​​of the distance between devices and the zero-sequence current with the theoretical reference value and noise variance to quantify the probability of the measured value occurring under a given fault state. This provides a mathematical basis for fault state judgment during information exchange between end-side neighbor nodes, improves the quantitative accuracy of preliminary diagnosis, and reduces collaborative positioning deviations caused by misjudgment of a single parameter.

[0035] Example 3

[0036] Combine Figure 3 This embodiment proposes a system architecture adapted to the above functions, which is built based on the cloud-native network fault diagnosis model library of Example 1 and the end-cloud-network collaborative positioning method of Example 2. This embodiment is a preferred application form of Examples 1 and 2, and its implementation does not affect the independent implementation of Examples 1 and 2.

[0037] The system architecture used to implement the above model library and collaborative positioning method includes: The device-side layer consists of terminal devices with data collection and preliminary diagnostic capabilities, monitoring network status and device parameters through built-in sensing modules; The edge layer deploys edge servers responsible for receiving end-side data, pre-processing, local fault analysis, and interacting with the end-side and cloud. The cloud layer includes cloud servers and a fault diagnosis model library. The cloud servers provide computing resources to support model training, version management, and global fault location. The model library stores optimized containerized models.

[0038] Specifically, the end-side layer is responsible for data collection and preliminary diagnosis, the edge layer processes data and performs local analysis, and the cloud layer provides computing resources and model libraries to support deep positioning. The three-layer architecture realizes layered collaboration of data processing and fault diagnosis. It not only uses the end-side and edge layers to reduce data transmission and cloud pressure, but also ensures positioning accuracy through global analysis of the cloud layer, adapting to the efficient operation and maintenance needs of complex networks.

[0039] The formula of the present invention is numerically calculated after dimensioning, and conventional means such as standardization can be used for dimensioning; the formula is generated based on a large amount of data simulation and is close to the actual situation. The preset parameters are set by those skilled in the art according to the scenario.

[0040] The present invention may be implemented through software, hardware, firmware, or a combination thereof. When implemented through software, it may be embodied as a computer program product whose instructions, when executed on a computer, implement the processes or functions of the present invention. The computer may be a general-purpose computer or a special-purpose computer, and the instructions may be stored on a readable medium or transmitted between media (including various transmission methods). Readable media include all types of computer-accessible media and related data storage devices, including magnetic, optical, and semiconductor media.

[0041] It should be noted that the program numbers do not represent the execution order, which is determined by the functional logic.

[0042] The division of units and systems in the embodiments is a logical division of functions, but other division methods are possible. The coupling and connection between units can be direct or indirect, including electrical, mechanical, etc. Functional units can be integrated, exist independently, or be integrated into multiple units.

[0043] If the functionality is implemented as a software unit and sold independently, it can be stored on a readable medium. The relevant software product contains instructions that enable a computer device to perform some or all of the steps of the method of the present invention. The medium includes a USB flash drive, a mobile hard drive, ROM, RAM, a magnetic disk, an optical disk, etc.

[0044] The above are only specific embodiments of the present application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A cloud-native network fault diagnosis model library, characterized by: include: The basic model construction module uses the orthogonal optimization design method combined with the convolutional neural network to build a basic fault diagnosis model. The model is trained using historical fault data and the model parameters are optimized through range analysis and variance analysis. When building the model, the modeling method is selected according to the fault type. The decision tree algorithm is used to model abnormal equipment settings, and the CNN algorithm is used to model abnormal network status. The input features are electrical quantity features. The model management and iteration module encapsulates the optimized model as a container image, manages the model version through the version control system, and records the training data and performance indicators of each iteration. Based on the cloud-native automated deployment mechanism, it supports grayscale release and dynamic updates of the model.

2. A method for collaborative positioning of an end-cloud network, using the cloud-native network fault diagnosis model library according to claim 1, characterized in that: The steps include: Step 1: Initial diagnosis and information exchange on the device side: The device-deployed perception module collects network status data and device operating parameters, uses a local lightweight model to make a preliminary fault diagnosis, compares the device operating parameters and network status data with preset fault thresholds, and marks the fault type. If the number of neighboring nodes is greater than or equal to a preset condition, it exchanges information with the neighboring nodes and calculates the likelihood function. Otherwise, it uses the Kalman filter algorithm to generate location information. Step 2: Narrowing the edge layer fault scope: Initial fault data is aggregated and preprocessed on the edge node receiving side. Combined with locally stored network topology information, the fault propagation model is used to narrow the fault scope. The fault suspicion of edge nodes is calculated using a fault propagation probability model. The edge nodes are ranked from high to low in terms of their suspected faults. The fault suspicion and propagation paths of a pre-set number of edge nodes / links are taken as the processing results and uploaded to the cloud. Step 3: Cloud-based in-depth analysis and fault location: The cloud calls the model in the fault diagnosis model library, performs in-depth analysis on the fault feature data uploaded by the edge node, locates the root cause of the fault through multi-model fusion reasoning, and generates results including the cause of the fault and the repair plan, which are fed back to the edge node and end-side device.

3. A system architecture for implementing the model library described in claim 1 and the collaborative positioning method described in claim 2, characterized in that: include: The device-side layer consists of terminal devices with data collection and preliminary diagnostic capabilities, monitoring network status and device parameters through built-in sensing modules; The edge layer deploys edge servers responsible for receiving end-side data, pre-processing, local fault analysis, and interacting with the end-side and cloud. The cloud layer includes cloud servers and a fault diagnosis model library. The cloud servers provide computing resources to support model training, version management, and global fault location. The model library stores optimized containerized models.

4. The cloud-native network fault diagnosis model library according to claim 1, characterized in that: The orthogonal optimization design method is specifically as follows: Determine the factors and levels, including convolution kernel size, learning rate, and batch size. Establish an orthogonal experimental parameter table to obtain model performance data for different parameter combinations. Use the range analysis method to calculate the performance mean and range of each factor at different levels. Use the variance analysis method to calculate the total sum of squares, factor sum of squares, and error sum of squares. Judge the significance of the factors through the mean square and significance evaluation value. Combine the range and variance analysis results to select the optimal level of each factor to form the optimal parameter combination.

5. The cloud-native network fault diagnosis model library according to claim 1, characterized in that: When building a basic fault diagnosis model, identify the fault type, including abnormal equipment setting values ​​and abnormal network status. If the fault type is abnormal equipment setting values, use the decision tree algorithm to build the model; if the fault type is abnormal network status, use the convolutional neural network algorithm to build the model. Identify electrical quantity characteristics from equipment operation data, including three-phase voltage RMS, current phase angle, current direct-axis / quadrature-axis components, and zero-sequence current; and use the electrical quantity characteristics as model input.

6. The device-cloud-network collaborative positioning method according to claim 2, characterized in that: The local lightweight model specifically includes a decision tree or a naive Bayesian model; the preliminary fault judgment is performed by comparing the equipment operating parameters with the preset fault thresholds corresponding to the parameters. When the value of the equipment operating parameter is greater than or equal to the preset fault threshold corresponding to the parameter, it indicates that the corresponding parameter has a potential fault, and the parameter is recorded as an equipment fixed value abnormality; At the same time, by comparing the parameters in the network status data with the preset fault thresholds corresponding to the parameters, when the parameters in the network status data are greater than or equal to the preset fault thresholds corresponding to the parameters, it indicates that there is a potential fault in the corresponding parameters, and the parameters are recorded as network status abnormalities.

7. The device-cloud-network collaborative positioning method according to claim 2, characterized in that: The multi-model fusion reasoning method for locating the root cause of the fault is as follows: Input the fault feature data uploaded by the edge node into the convolutional neural network, XGBoost and Bayesian network; the fault feature data includes electrical quantity features and topological correlation features; The convolutional neural network extracts characteristic nonlinear patterns through multi-layer convolution kernels and outputs the first fault probability; XGBoost outputs the second failure probability based on gradient boosting decision tree ensemble learning; The Bayesian network infers the probability of fault conditions through a probabilistic graphical model and outputs the third fault probability; The weight optimization algorithm that minimizes the fusion error is used to determine the weight of each model. The first to third fault probabilities are fused through the weighted sum formula to obtain the fault judgment probability. .

8. The device-cloud-network collaborative positioning method according to claim 2, characterized in that: The likelihood function is used to describe the probability of occurrence of measurement values ​​under a given fault state, wherein the measurement values ​​include the distance measurement value between devices and the zero-sequence current value. The likelihood function adopts a Gaussian distribution model, whose mean is the theoretical reference value of the measurement value and the variance is the variance of the measurement noise.

Images