A manifold learning belief rule base (ML-BRB) based software-defined packet transport network (SPTN) explainable fault diagnosis method
By using a confidence rule base method based on manifold learning, the problems of rule explosion and feature semantic loss caused by high-dimensional monitoring data in SPTN networks are solved, and efficient and interpretable fault diagnosis is achieved.
Patent Information
- Application Number
- CN202610575391.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-28
- Publication Date
- 2026-07-03
AI Technical Summary
When faced with high-dimensional monitoring data from Software-Defined Packet Transport Networks (SPTN), existing technologies suffer from an exponential increase in the total number of rules in the Confidence Rule Base (BRB), and the loss of physical semantics of features after dimensionality reduction, resulting in uninterpretable models.
The Confidence Rule Base (ML-BRB) method based on manifold learning is adopted. The high-dimensional monitoring data is mapped to the low-dimensional manifold space through the t-SNE algorithm, the initial parameters of the confidence rule base model are adaptively determined, and the model parameters are optimized by using the projection covariance matrix to maintain the interpretability of the diagnostic conclusions.
It effectively solves the problem of rule explosion caused by high-dimensional monitoring data, while maintaining the physical semantics of features after dimensionality reduction, thus achieving interpretability and high accuracy in fault diagnosis.
Smart Images
Figure CN122339943A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of communication network fault diagnosis, and specifically to an interpretable fault diagnosis method for software-defined packet transport networks (SPTN) based on a manifold learning based belief rule base (ML-BRB). Background Technology
[0002] Modern communication networks, such as Software-Defined Packet Transport Networks (SPTN), serve as critical infrastructure and are subject to stringent Service Level Agreements (SLAs) constraints on fault recovery time. Due to the deep coupling between the network application layer, control layer, and forwarding layer, faults often exhibit significant cross-layer propagation characteristics, resulting in highly dimensional and nonlinear monitoring data.
[0003] Existing fault diagnosis methods are mainly divided into three categories: The first category is purely qualitative methods, which rely on expert experience to determine faults. These methods are highly subjective and difficult to handle complex cross-layer faults. The second category is quantitative data-driven methods, such as deep learning and support vector machines. Although they perform well in terms of accuracy, they are essentially "black box" models. Their decision-making logic cannot be directly understood by maintenance personnel, making it difficult to meet the stringent interpretability requirements of high-risk decision-making scenarios. The third category is the Belief Rule Base (BRB) method. As a semi-quantitative knowledge representation and reasoning framework, it embeds expert knowledge through the "IF-THEN" rule structure and uses the Evidence Reasoning (ER) algorithm to fuse fuzzy and incomplete information. It has become an interpretable fault diagnosis method that balances accuracy and transparency.
[0004] However, BRB (Browser Logic Builder) encounters a "combinatorial explosion" dilemma when dealing with high-dimensional network monitoring data: the total number of rules grows exponentially with the dimension of the premise attributes. When facing high-dimensional monitoring indicators, even with only a few reference values per dimension, the total number of rules expands exponentially, far exceeding the physical processing capacity. To alleviate this problem, existing research has proposed methods such as decomposition based on directed acyclic graphs (DAGs), rule reduction, data-driven parameter determination, and feature dimensionality reduction. However, DAG decomposition does not fundamentally reduce the input dimension; rule reduction in high-dimensional space leads to distortion of similarity measurement due to the "curse of dimensionality"; data-driven methods are limited to the original high-dimensional input space, relying on membership functions to characterize matching relationships, resulting in unreliable statistical results under strongly coupled data; and existing dimensionality reduction methods are mostly limited to linear techniques such as principal component analysis (PCA), lacking the ability to preserve local manifold structures.
[0005] More importantly, while dimensionality reduction techniques such as manifold learning can alleviate the curse of dimensionality, they cause low-dimensional features to lose their original physical semantics. For BRB, the core of its semi-quantitative nature lies in the fact that the reference values of the premise attributes are usually directly specified by experts based on clear physical meaning, thereby embedding prior knowledge. If the low-dimensional features are merely a set of abstract numerical symbols, experts cannot provide meaningful initial reference values and confidence levels, leading BRB into the dilemma of "rules that cannot be interpreted after dimensionality reduction".
[0006] Therefore, there is an urgent need for a general fault diagnosis method that can maintain the interpretability of BRB while reducing dimensions and decoupling, and can trace the diagnostic conclusions back to the original monitoring data. Summary of the Invention
[0007] This invention provides an interpretable fault diagnosis method for SPTN networks based on ML-BRB, aiming to solve the problems of high-dimensional coupling of monitoring indicators in SPTN networks leading to an explosion of confidence rule base combinations, and the loss of physical semantics of low-dimensional features after manifold dimensionality reduction, resulting in the inability to initialize parameters.
[0008] The present invention is implemented as follows: an interpretable fault diagnosis method for SPTN networks based on ML-BRB includes the following steps:
[0009] Collect multidimensional monitoring data from the SPTN network, construct a high-dimensional monitoring indicator system, and perform data preprocessing;
[0010] The SPTN network state manifold space hypothesis is proposed, and the t-SNE algorithm is used to map the preprocessed high-dimensional monitoring data to the low-dimensional manifold space to obtain the low-dimensional manifold coordinates of each sample.
[0011] Based on manifold geometry and dynamic neighborhood statistics, the initial parameters of the confidence rule base model are adaptively determined. The initial parameters include reference values of the premise attributes and the confidence distribution of the rules.
[0012] The reference values of each dimension are combined by Cartesian product to generate a confidence rule base, and the model parameters are optimized by an adaptive evolution strategy of projection covariance matrix.
[0013] By using the IF-THEN rule structure of the confidence rule base and the evidence reasoning algorithm, the fault type and confidence distribution with physical meaning are output.
[0014] By utilizing the one-to-one correspondence between samples before and after dimensionality reduction, diagnostic conclusions can be traced back to the original high-dimensional monitoring data, providing a data foundation for maintenance decisions.
[0015] Furthermore, the data preprocessing includes:
[0016] Delete the abnormal column containing an invalid index;
[0017] Missing values are filled using the sample median, and infinity values are replaced with the maximum feature value.
[0018] Delete the constant column with variance of 0;
[0019] A quantile truncation strategy is used to handle outliers;
[0020] The truncated data is standardized to ensure that each feature has zero mean and unit variance.
[0021] Furthermore, the network state manifold space is assumed to be:
[0022] Although the original observation space is high-dimensional, the system's operating state is essentially embedded in a low-dimensional smooth manifold. That is, there exists a manifold structure with an intrinsic dimension much lower than the observation dimension, making all high-dimensional observation data points nonlinear mappings of this low-dimensional manifold.
[0023] The hypothesis is verified by the following geometric topology criterion:
[0024] Intraclass compactness: Samples of the same type of working condition maintain local neighborhood relationships on the low-dimensional manifold, exhibiting a compact clustered distribution;
[0025] Inter-class discreteness: Sample clusters under different operating conditions exhibit significant topological separation on the manifold, forming clear decision boundaries;
[0026] Evolutionary continuity: For gradually changing characteristics, the sample points exhibit a continuous spatial trajectory or regional transition on the manifold.
[0027] Furthermore, the step of mapping the preprocessed high-dimensional monitoring data to a low-dimensional manifold space using the t-SNE algorithm includes:
[0028] Calculate the conditional probabilities between data points in a high-dimensional space and construct a symmetric joint probability distribution;
[0029] Randomly initialize mapping points in the low-dimensional target space;
[0030] The joint probability of point pairs in low-dimensional space is calculated using the t-distribution;
[0031] By minimizing the KL divergence between the high-dimensional and low-dimensional distributions, the embedding coordinates are optimized using gradient descent with momentum.
[0032] Repeatedly calculate the low-dimensional joint probability and gradient descent optimization until the maximum number of iterations is reached or the loss function converges.
[0033] Furthermore, the initial parameters of the confidence rule base model adaptively determined based on manifold geometry and dynamic neighborhood statistics include:
[0034] Adaptive determination of prerequisite attribute reference values:
[0035] Calculate the coordinates of the center point in the low-dimensional manifold space for each working condition;
[0036] Generate a preset number of initial reference values at equal intervals within the range of values for each dimension;
[0037] Arrange the center points of each working condition in ascending order, and calculate the midpoint of adjacent center points as a candidate set of fault mutation points;
[0038] The initial reference value is adjusted toward the nearest fault mutation point, and the adjustment range is controlled by the relaxation factor, while the first and last reference values remain unchanged.
[0039] Confidence estimation based on dynamic neighborhood radius:
[0040] For each reference value in each dimension, the dynamic neighborhood radius is determined based on the distance between its adjacent reference values;
[0041] Determine the neighborhood sample set, that is, all samples whose coordinates in this dimension fall within the neighborhood radius interval;
[0042] Estimate the confidence level, which is the ratio of the number of samples belonging to each type of fault in the neighborhood to the total number of samples in the neighborhood. If the neighborhood sample set is empty, the nearest neighbor rule is used to assign the value.
[0043] Furthermore, the dynamic neighborhood radius is adaptively determined based on the local spacing of the reference value:
[0044] In regions with dense reference values, the neighborhood radius is small;
[0045] In regions where reference values are sparse, the neighborhood radius is larger;
[0046] This allows for adaptive matching of the local distribution density of the data.
[0047] Furthermore, the generation of the confidence rule base includes:
[0048] The reference values of the premise properties of each manifold are combined by Cartesian product to generate a preset number of confidence rules;
[0049] The consequent confidence distribution of each rule is obtained by fusing the confidence scores estimated independently for each dimension using an evidence reasoning parsing algorithm.
[0050] Furthermore, the optimization of model parameters using the adaptive evolution strategy of the projection covariance matrix includes:
[0051] The objective function is the mean absolute error on the training set.
[0052] The search direction is adaptively updated by the covariance matrix, and the projection operation is used to ensure that the parameters meet the normalization constraints.
[0053] Iterative optimization continues until the loss function converges, yielding the optimal parameter set.
[0054] Furthermore, the method of using the one-to-one correspondence between samples before and after dimensionality reduction to trace the diagnostic conclusion back to the original high-dimensional monitoring data includes:
[0055] As a sample-by-sample mapping algorithm, t-SNE maintains a strict one-to-one correspondence between the reduced low-dimensional coordinates and the original high-dimensional samples through the sample index.
[0056] When the confidence rule base diagnoses a sample as a specific fault, it directly locates the complete indicator vector of the sample in the original high-dimensional monitoring data through the sample index.
[0057] Based on the fault type conclusion with physical semantics, maintenance personnel can directly retrieve the original high-dimensional monitoring data of the sample and formulate maintenance strategies.
[0058] This invention also provides an interpretable fault diagnosis system for SPTN networks using ML-BRB, comprising:
[0059] The data acquisition and preprocessing module is used to acquire multidimensional monitoring data of the target network and perform data preprocessing.
[0060] The manifold dimensionality reduction module is used to propose the network state manifold space hypothesis and use the t-SNE algorithm to map the preprocessed high-dimensional monitoring data to the low-dimensional manifold space.
[0061] The parameter initialization module is used to adaptively determine the initial parameters of the confidence rule base model based on manifold geometry and dynamic neighborhood statistics.
[0062] The rule base construction and optimization module is used to combine reference values of various dimensions by Cartesian product to generate a confidence rule base, and to optimize the model parameters by using an adaptive evolution strategy of projection covariance matrix.
[0063] The fault diagnosis module is used to output fault types and confidence distributions with physical meaning by using the IF-THEN rule structure of the confidence rule base and the evidence reasoning algorithm.
[0064] The data backtracking module is used to leverage the one-to-one correspondence between samples before and after dimensionality reduction to trace diagnostic conclusions back to the original high-dimensional monitoring data. Attached Figure Description
[0065] Figure 1 This is a flowchart illustrating the network-interpretable fault diagnosis method based on a manifold learning belief rule base provided by the present invention.
[0066] Figure 2 This is a schematic diagram of the SPTN network architecture provided by the present invention;
[0067] Figure 3 The schematic diagram of adaptive adjustment of prerequisite attribute reference values and adaptive dynamic neighborhood radius based on manifold geometry provided by this invention; Detailed Implementation
[0068] The technical solutions of 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0069] This invention provides a network interpretable fault diagnosis method based on ML-BRB. This method directly addresses two core challenges in fault diagnosis of modern communication networks (such as SPTN networks): First, high-dimensional monitoring indicators trigger a "combinatorial explosion" in the confidence rule base (BRB), making model construction impossible; second, while conventional dimensionality reduction methods can alleviate the curse of dimensionality, they cause low-dimensional features to lose their original physical semantics, preventing the BRB from relying on expert knowledge for meaningful parameter initialization, thus losing the foundation of its interpretability.
[0070] To overcome the aforementioned challenges simultaneously, the method of this invention follows a technical route of "hypothesis guidance—structural decoupling—semantic reconstruction—conclusion backtracking." First, the "network state manifold space hypothesis" is proposed and verified, providing a theoretical basis for dimensionality reduction using manifold learning, fundamentally compressing the input space of BRB. Second, addressing the new dilemma that coordinates after dimensionality reduction are merely abstract geometric quantities without physical meaning, a parameter adaptive initialization mechanism based on "manifold geometry and dynamic neighborhood statistics" is designed. By mining the topological structure and statistical regularities of data in the low-dimensional manifold space, each reference value is assigned a confidence level with frequency interpretation, thereby reconstructing the interpretable semantics of BRB rules based on abstract features. Finally, utilizing the one-to-one correspondence of samples retained during dimensionality reduction, the diagnostic conclusions are backtracked to the original high-dimensional data, forming a complete interpretable diagnostic closed loop.
[0071] Figure 1 This is a flowchart illustrating the method provided by the present invention. (In conjunction with...) Figure 1 The method specifically includes the following steps:
[0072] Step S1: Collect multidimensional monitoring data of the target network, construct a high-dimensional monitoring indicator system, and perform data preprocessing.
[0073] In this embodiment, a Software-Defined Packet Transport Network (SPTN) is used as the object. Through southbound interface protocols (such as OpenFlow and gRPC), data on device performance and protocol operation status are collected from the application layer, control layer, and forwarding layer of the network at a certain sampling period, constructing a high-dimensional dataset containing high-dimensional monitoring indicators. The collected indicators cover multiple dimensions, including network performance, control module status, forwarding layer interface and optical module performance, and multi-protocol operation status.
[0074] To make high-dimensional data suitable for subsequent manifold learning analysis, it needs to be rigorously preprocessed. This preprocessing process constitutes an integral part of the technical solution of this invention, and its steps are as follows:
[0075] S101: Abnormal Column Cleanup. Automatically iterates through the column names of the dataset, deleting columns containing invalid index information such as "Unnamed", ensuring that all columns are valid monitoring features.
[0076] S102: Handling missing and infinite values. For missing values in the data, the median of the sample for that feature is used to fill them in, as shown in formula (1); for a few infinite values in the data caused by calculation or collection errors, they are replaced with the maximum value of that feature.
[0077]
[0078] S103: Constant column removal. Calculate the variance of each feature dimension and remove constant columns with zero variance. These indicators, which do not change with the system state, do not contain any diagnostic information; removing them reduces redundancy.
[0079] S104: Outlier Truncation. To reduce the interference of extreme outliers on subsequent data analysis, a quantile truncation strategy is adopted. For each feature dimension, its 1st percentile is calculated. and the 99th percentile and less than And greater than The values are respectively limited to the boundary.
[0080] S105: Z-score standardization. To eliminate the analytical bias caused by the different dimensions and value ranges of different indicators, the truncated data is subjected to Z-score standardization, as shown in formula (2), so that each dimension of the feature after processing has zero mean and unit variance.
[0081]
[0082] In the formula, Let be the mean of the j-th feature. Let be the standard deviation of the j-th feature.
[0083] Step S2: Propose the network state manifold space hypothesis, use the t-SNE algorithm to map the preprocessed high-dimensional monitoring data to the low-dimensional manifold space, and obtain the one-to-one correspondence between samples before and after dimensionality reduction.
[0084] This is a core step in addressing the BRB "combination explosion" problem and also the starting point for rebuilding interpretability.
[0085] S201: Propose the Network State Manifold Space Hypothesis. Based on a profound understanding of the operating mechanism of network systems, this invention proposes a preliminary technical hypothesis: although the original observation space has a dimension as high as 65, all operating states of the system (including normal states and various fault states) are essentially driven by a few underlying physical variables (such as resource exhaustion, physical link interruption, and protocol state machine transitions). Therefore, these high-dimensional observation points are not randomly scattered in space, but are embedded in a low-dimensional smooth manifold. Above this, the "network state manifold space hypothesis" forms the theoretical foundation for all subsequent technical aspects. Its validity can be verified using the following geometric topological criteria:
[0086] Intraclass compactness: Samples of the same operating conditions should exhibit a compact clustered distribution in the dimensionality-reduced low-dimensional space.
[0087] Inter-class discreteness: Sample clusters under different operating conditions should have significant topological separation in low-dimensional space, forming clear decision boundaries.
[0088] Evolutionary continuity: For gradual faults, the sample points should exhibit a continuous spatial trajectory on the manifold.
[0089] S202: Manifold Dimensionality Reduction Using t-SNE. To verify the above assumptions and extract low-dimensional features, this invention uses the t-distributed random neighborhood embedding (t-SNE) algorithm for dimensionality reduction. Compared to methods such as PCA and KPCA, the optimization objective of t-SNE focuses on preserving the local manifold structure of the data, perfectly matching the assumptions of this invention. The specific dimensionality reduction process is as follows:
[0090] Calculating high-dimensional joint probabilities: In a high-dimensional space, construct a Gaussian distribution centered at each point and calculate the conditional probabilities between data points. And by symmetry processing, a joint probability distribution P is constructed, as shown in formulas (3) and (4), to characterize the probability neighborhood structure on the high-dimensional manifold.
[0091] Initialize low-dimensional mapping: Randomly initialize the coordinates of mapping points in the low-dimensional target space (3-dimensional in this embodiment). .
[0092] Calculating the joint probability in low-dimensional space: In low-dimensional space, the joint probability of point pairs is calculated using a heavy-tailed t-distribution. Construct the distribution Q as shown in formula (5). This step uses the t-distribution to solve the "crowding problem" between different manifold branches.
[0093] Gradient descent optimization: The objective function is to minimize the KL divergence C between the high-dimensional distribution P and the low-dimensional distribution Q, as shown in Equation (6), and to perform iterative optimization according to the gradient shown in Equation (7) until convergence or the preset maximum number of iterations is reached.
[0094]
[0095] Obtaining low-dimensional manifold coordinates and their correspondences: After iteration, for each original high-dimensional sample... Obtain its unique low-dimensional manifold coordinates It is important to emphasize that t-SNE, as a sample-by-sample mapping algorithm, strictly maintains the one-to-one correspondence between sample indices during its dimensionality reduction process. This property is crucial for the subsequent implementation of "sample backtracking".
[0096] After this step, the input dimension of the BRB model drops dramatically from high dimension to low dimension, and the total number of rules explodes exponentially (e.g., ...). This reduces the problem to a physically solvable level, fundamentally solving the combinatorial explosion problem. However, coordinates... At this point, it is merely a set of abstract numbers without any physical meaning.
[0097] Step S3: Based on manifold geometry and dynamic neighborhood statistics, adaptively determine the initial parameters of the confidence rule base model, and reconstruct the interpretability of the rules on top of the abstract features.
[0098] This is a crucial step in solving the core problem of "features losing their physical semantics after dimensionality reduction." This invention designs a completely data-driven parameter initialization mechanism. Its core idea is to define each reference value and its corresponding confidence level by analyzing the geometric distribution of data points on the manifold after dimensionality reduction and the statistical characteristics of neighboring samples, rather than relying on expert interpretation of the physical meaning of abstract coordinates. This allows the consequences of each rule to be traced back to the original samples, giving them an objective frequency interpretation.
[0099] S301: Adaptive Determination of Prerequisite Attribute Reference Values. This step aims to adjust the initially equidistant reference values to values that better reflect the evolution boundaries between different fault conditions. Specifically, it includes:
[0100] Calculation of Operating Conditions Center: For each type of fault condition... Calculate its centroid (center point coordinates) in the low-dimensional manifold space, as shown in formula (10).
[0101] Generate initial reference values: Within the data range of each dimension, generate H initial reference values at equal intervals. As shown in formula (11).
[0102] Extracting the candidate set of fault mutation points: Sort the operating condition center points in ascending order for each dimension, and calculate the midpoint between every two adjacent center points. As shown in Equation (12), these midpoints geometrically represent the potential boundary locations of different failure modes on the manifold.
[0103] Reference value adaptive adjustment: Adjust each internal initial reference value Towards the nearest fault mutation point The direction is adjusted using attraction to obtain the final reference value. As shown in formula (13). Here, a relaxation factor is introduced. Control the adjustment intensity, keeping the initial and final reference values unchanged. Figure 3 This illustrates the adjustment process.
[0104] S302: Confidence estimation based on dynamic neighborhood radius. For each reference value... Assignment to various types of faults initial confidence level The process is entirely data-driven, objective, and interpretable.
[0105] Determine the dynamic neighborhood radius To adapt to the density differences of data in different regions, the neighborhood radius is adaptively calculated based on the distance between the current reference value and its adjacent reference values, as shown in formula (14). This is the neighborhood radius scaling factor. In regions with dense reference values (dense data), the neighborhood radius automatically shrinks; while in regions with sparse reference values, the neighborhood radius automatically expands.
[0106] Define the neighborhood sample set Collect all pairs of coordinates whose d-th dimension falls within the interval [0, 1]. The samples within the range constitute the neighborhood sample set of the reference value.
[0107] Estimating confidence level: Statistical neighborhood sample set The middle belongs to various types of faults The sample proportion is directly used as the initial confidence level for various types of faults. As shown in formula (15). If the neighborhood is empty, the nearest neighbor rule is used for assignment. Through the above steps, this invention achieves a key innovation: even if the input features are abstract low-dimensional coordinates, the meaning of the BRB rules remains clear. Step S4: Generate an initial confidence rule base and optimize the model parameters using an adaptive evolution strategy based on the projection covariance matrix.
[0108] S401: Construct the initial BRB. Combine the H reference values of each of the three manifold premise properties using a Cartesian product to generate a total of... There are 10 rules. The consequent confidence distribution of each rule is obtained by fusing the independent confidence scores of the reference values corresponding to the three dimensions constituting the premise of the rule using the Evidence Reasoning (ER) analytical algorithm. The initial rule weights and attribute weights are both set to 1. The initial model constructed in this step already possesses considerable fault differentiation capability and interpretability.
[0109] S402: Parameter Optimization. To further improve diagnostic accuracy, this invention uses the mean absolute error (MAE) on the training set as the objective function. As shown in Equation (16), the Projected Covariance Matrix Adaptive Evolution Strategy (P-CMA-ES) is used to fine-tune and optimize the rule weights, attribute weights, and confidence scores of the BRB. The projection operation ensures that the parameters always satisfy the normalization constraints, as shown in Equation (17). As an efficient optimizer, P-CMA-ES can effectively search for the global optimum in this high-dimensional, constrained parameter space and adjust the decision boundary.
[0110]
[0111] Step S5: Using the IF-THEN rule structure of the confidence rule base and the evidence reasoning algorithm, output the fault type and confidence distribution with physical meaning.
[0112] When applying the model, for a new sample to be diagnosed, the same preprocessing and dimensionality reduction as steps S1 and S2 are first performed to obtain its manifold coordinates. Then, the matching degree of these coordinates with the premise attributes of each rule is calculated, and then the activation weight of each rule is calculated. Finally, using the Evidence Reasoning (ER) analytical algorithm, as shown in formulas (18) and (19), the consequent confidence distributions of all activated rules are fused to obtain the final diagnostic result. .
[0113] The output not only provides a decision on "what kind of fault it is", but also provides a confidence level distribution on "how likely it is", providing key information for operations and maintenance personnel to assess uncertainty.
[0114] Step S6: Utilize the one-to-one correspondence between samples before and after dimensionality reduction to trace the diagnostic conclusions back to the original high-dimensional monitoring data, providing a data foundation for maintenance decisions.
[0115] This is the final step in achieving an interpretable decision-making closed loop in this invention. When the BRB diagnoses a specific fault in a low-dimensional manifold coordinate, maintenance personnel can not only see the reasoning logic and confidence level, but also trigger a crucial action: using the one-to-one correspondence of the sample indexes retained by t-SNE in step S2, they can directly retrieve the original high-dimensional monitoring data corresponding to the current low-dimensional coordinate from the database. This allows maintenance personnel to directly address indicators with clear physical meanings, such as "port down," "low optical power," and "high CPU utilization," and, combined with the conclusion of the fault type, quickly formulate and execute precise emergency repair strategies (such as replacing optical modules, restarting protocol processes, and load balancing switching).
[0116] Thus, this invention has fully realized a diagnostic closed loop of "manifold decoupling - rule interpretation - sample backtracking", which, while ensuring high accuracy and real-time performance of diagnosis, gives the entire fault diagnosis process unprecedented transparency and traceability.
[0117] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A manifold learning belief rule base (ML-BRB) based software defined packet transport network (SPTN) explainable fault diagnosis method, characterized in that, Includes the following steps: Step 1: Collect multidimensional monitoring data of the target network, construct a high-dimensional monitoring indicator system, and perform data preprocessing; The network state manifold space hypothesis is proposed, and the t-SNE algorithm is used to map the preprocessed high-dimensional monitoring data to the low-dimensional manifold space to obtain the low-dimensional manifold coordinates of each sample. Step 2: Based on manifold geometry and dynamic neighborhood statistics, the initial parameters of the confidence rule base model are adaptively determined. The initial parameters include reference values of the premise attributes and the confidence distribution of the rules. The reference values of each dimension are combined by Cartesian product to generate the confidence rule base, and the model parameters are optimized by an adaptive evolution strategy of projection covariance matrix. Step 3: Using the IF-THEN rule structure and evidence reasoning algorithm of the confidence rule base, output the fault type and confidence distribution with physical meaning; using the one-to-one correspondence between samples before and after dimensionality reduction, trace the diagnostic conclusion back to the original high-dimensional monitoring data to provide a data basis for maintenance decisions.
2. The method according to claim 1, characterized in that, The data preprocessing includes: Delete the abnormal column containing an invalid index; Missing values are filled using the sample median, and infinity values are replaced with the maximum feature value. Delete the constant column with variance of 0; A quantile truncation strategy is used to handle outliers; The truncated data is standardized to ensure that each feature has zero mean and unit variance.
3. The network state manifold space assumption according to claim 1 is as follows: Although the original observation space is high-dimensional, the system's operating state is essentially embedded in a low-dimensional smooth manifold. That is, there exists a manifold structure with an intrinsic dimension much lower than the observation dimension, making all high-dimensional observation data points nonlinear mappings of this low-dimensional manifold. The hypothesis is verified using three geometric topological criteria: intra-class compactness, inter-class discreteness, and evolutionary continuity.
4. Mapping preprocessed high-dimensional monitoring data to a low-dimensional manifold space using the t-SNE algorithm as described in claim 1, comprising: Calculate the conditional probabilities between data points in a high-dimensional space and construct a symmetric joint probability distribution; Randomly initialize mapping points in the low-dimensional target space; The joint probability of point pairs in low-dimensional space is calculated using the t-distribution; By minimizing the KL divergence between the high-dimensional and low-dimensional distributions, the embedding coordinates are optimized using gradient descent with momentum. Repeatedly calculate the low-dimensional joint probability and gradient descent optimization until the maximum number of iterations is reached or the loss function converges.
5. The method according to claim 1, characterized in that, The initial parameters of the confidence rule base model, which is adaptively determined based on manifold geometry and dynamic neighborhood statistics, include: (1) Adaptive determination of prerequisite attribute reference values: Calculate the coordinates of the center point in the low-dimensional manifold space for each working condition; Generate a preset number of initial reference values at equal intervals within the range of values for each dimension; Arrange the center points of each working condition in ascending order, and calculate the midpoint of adjacent center points as a candidate set of fault mutation points; The initial reference value is adjusted toward the nearest fault mutation point, and the adjustment range is controlled by the relaxation factor, while the first and last reference values remain unchanged. (2) Confidence estimation based on dynamic neighborhood radius: For each reference value in each dimension, the dynamic neighborhood radius is determined based on the distance between its adjacent reference values; Determine the neighborhood sample set, that is, all samples whose coordinates in this dimension fall within the neighborhood radius interval; Estimate the confidence level, which is the ratio of the number of samples belonging to each type of fault in the neighborhood to the total number of samples in the neighborhood. If the neighborhood sample set is empty, the nearest neighbor rule is used to assign the value.
6. The method according to claim 5, characterized in that, The dynamic neighborhood radius is adaptively determined based on the local spacing of the reference value: In regions with dense reference values, the neighborhood radius is small; In regions where reference values are sparse, the neighborhood radius is larger; This allows for adaptive matching of the local distribution density of the data.
7. The method according to claim 1, characterized in that, The confidence rule base includes: The reference values of the premise properties of each manifold are combined by Cartesian product to generate a preset number of confidence rules; The consequent confidence distribution of each rule is obtained by fusing the confidence scores estimated independently for each dimension using an evidence reasoning parsing algorithm.
8. The method according to claim 1, characterized in that, The optimization of model parameters using the adaptive evolution strategy of projection covariance matrix includes: The objective function is the mean absolute error on the training set. The search direction is adaptively updated by the covariance matrix, and the projection operation is used to ensure that the parameters meet the normalization constraints. Iterative optimization continues until the loss function converges, yielding the optimal parameter set.
9. The method according to claim 1, characterized in that, The method of using the one-to-one correspondence between samples before and after dimensionality reduction to trace diagnostic conclusions back to the original high-dimensional monitoring data includes: As a sample-by-sample mapping algorithm, t-SNE maintains a strict one-to-one correspondence between the reduced low-dimensional coordinates and the original high-dimensional samples through the sample index. When the confidence rule base diagnoses a sample as a specific fault, it directly locates the complete indicator vector of the sample in the original high-dimensional monitoring data through the sample index. Based on the fault type conclusion with physical semantics, maintenance personnel can directly retrieve the original high-dimensional monitoring data of the sample and formulate maintenance strategies.
10. An interpretable fault diagnosis system for SPTN networks based on ML-BRB, characterized in that, include: The data acquisition and preprocessing module is used to acquire multidimensional monitoring data of the target network and perform data preprocessing. The manifold dimensionality reduction module is used to propose the network state manifold space hypothesis and use the t-SNE algorithm to map the preprocessed high-dimensional monitoring data to the low-dimensional manifold space. The parameter initialization module is used to adaptively determine the initial parameters of the confidence rule base model based on manifold geometry and dynamic neighborhood statistics. The rule base construction and optimization module is used to combine reference values of various dimensions by Cartesian product to generate a confidence rule base, and to optimize the model parameters by using an adaptive evolution strategy of projection covariance matrix. The fault diagnosis module is used to output fault types and confidence distributions with physical meaning by using the IF-THEN rule structure of the confidence rule base and the evidence reasoning algorithm. The data backtracking module is used to leverage the one-to-one correspondence between samples before and after dimensionality reduction to trace diagnostic conclusions back to the original high-dimensional monitoring data.