A Fault Prediction Method and System for Optical Networks Based on Multimodal Dynamic Fusion

Abstract

This invention relates to a method and system for optical network fault prediction based on multimodal dynamic fusion, belonging to the field of optical communication network operation and maintenance technology. This invention solves the problems of poor adaptability and unreliable prediction results caused by the lack of physical constraints in existing technologies that use fixed weights to fuse multimodal data. This invention simultaneously collects key performance indicators, log text, and network topology flow, and extracts temporal, semantic, and spatial features in parallel. Through a knowledge-guided cross-attention mechanism, it utilizes an optical network knowledge graph to guide the adaptive fusion of multimodal features and introduces modal confidence evaluation to achieve dynamic weighting. Finally, it combines physical consistency loss to verify the output fault probability. This invention effectively improves the accuracy, robustness, and interpretability of optical network fault prediction, ensuring that the prediction results conform to physical laws.

Description

Technical Field

[0001] This invention belongs to the field of optical communication network operation and maintenance technology, and relates to a method and system for optical network fault prediction based on multimodal dynamic fusion. Background Technology

[0002] With the explosive growth of cloud computing, big data, and artificial intelligence, data centers, as the foundation of the digital economy, are experiencing a continuous increase in scale and complexity. Optical networks, with their high bandwidth, low latency, and high reliability, have become the core infrastructure carrying traffic within and between data centers. However, core components of optical networks, such as optical modules, fiber optic links, and switching equipment, are susceptible to various failures during long-term operation due to factors such as temperature fluctuations, signal attenuation, electromagnetic interference, and hardware aging. These failures are characterized by their high degree of concealment, rapid propagation, and wide-ranging impact; once they occur, they can lead to large-scale service interruptions, causing significant economic losses and social impact. Therefore, building an efficient and accurate optical network failure prediction mechanism is crucial for ensuring the stable operation of data centers.

[0003] In the field of optical network fault prediction, traditional solutions often employ a single data source for analysis. For example, they rely solely on time-series data of key performance indicators (KPIs) such as optical module transmit and receive power, operating voltage, bias current, and operating temperature for anomaly detection. However, a single indicator often fails to comprehensively reflect the complex network conditions, leading to both false positives and false negatives. To improve prediction accuracy, related technologies have begun to explore the introduction of multi-source data for fusion analysis. Common data modalities include KPI time-series data reflecting the physical state of equipment, log text data recording system operating status and abnormal events, and network topology data demonstrating device connectivity and physical architecture.

[0004] Existing multimodal fault prediction methods typically employ early or mid-stage fusion strategies, directly concatenating feature vectors from different modalities at a specific level or superimposing them with fixed weights. These methods have significant drawbacks. Firstly, the quality of data from different modalities varies and changes dynamically over time. For instance, during normal equipment operation, log text often contains meaningless routine information with low reference value; however, during the fault latency period, KPI data may fluctuate slightly, at which point error messages in the logs have extremely high diagnostic value. Fixed-weight fusion methods cannot adaptively adjust to the real-time information quality of each modality, resulting in the suppression of effective information and the amplification of noise, severely impacting the model's diagnostic accuracy in complex scenarios.

[0005] On the other hand, existing models are mostly built on purely data-driven machine learning frameworks, lacking constraints on the physical laws governing optical communication. The occurrence and evolution of optical network faults follow specific physical laws and engineering logic; for example, optical power attenuation inevitably leads to an increase in the bit error rate, and a specific link interruption inevitably affects downstream nodes. Purely data-driven models, when lacking sufficient training samples or with skewed data distribution, are highly prone to producing predictions that defy common sense, not only reducing the reliability of fault diagnosis but also making the model's decision-making process lack interpretability and difficult for maintenance personnel to accept.

[0006] Furthermore, the operation and maintenance environment of optical networks is complex, often facing situations where data collection is incomplete or data contamination occurs. Traditional fusion models often experience a sharp decline in performance when data for a particular modality is missing, lacking sufficient robustness and fault tolerance. Therefore, how to effectively fuse multimodal heterogeneous data, achieve dynamic perception and adaptive weighting of data quality in the model, and deeply integrate the physical laws of optical communication into the model training and inference process has become a key issue that current optical network fault prediction technology urgently needs to address. Summary of the Invention

[0007] In view of this, the purpose of the present invention is to provide a method and system for optical network fault prediction based on multimodal dynamic fusion.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A fault prediction method for optical networks based on multimodal dynamic fusion includes the following steps: Data input steps: Synchronously collect key performance indicators of the optical network, namely KPI data stream, log text stream, and network topology stream, and preprocess the three types of data streams; Feature representation steps: Extract the temporal feature vector of the KPI data stream using a bidirectional long short-term memory network, i.e., Bi-LSTM. The semantic feature vector of the log text stream is extracted using a BERT pre-trained model. The spatial feature vector of the network topology flow is extracted using a graph convolutional network, i.e., a graph network convolutional network (GCN). ; Knowledge fusion steps: Construct an optical network knowledge graph to obtain knowledge vectors Based on the knowledge vector For the time-series feature vector semantic feature vector and spatial eigenvectors Knowledge-oriented cross-attention calculation is performed, and dynamic weighted fusion is combined with modality confidence scores to obtain a multimodal fusion vector. ; Fault prediction step: The multimodal fusion vector The input is a multilayer perceptron (MLP) model, and the failure probability distribution is output through the Softmax function. At the same time, physical consistency loss is introduced to verify the prediction results.

[0009] Furthermore, the data input step specifically includes: The KPI data stream is repaired by linear interpolation and then normalized to zero mean. The log text stream is subjected to a fixed-depth tree parsing algorithm for log template matching, followed by word segmentation, cleaning, and stop word filtering. The network topology flow relies on the link layer discovery protocol to detect the interconnection relationship and physical architecture of device nodes in real time.

[0010] Furthermore, in the feature representation step, the temporal feature vector From forward hidden state and backward hidden state It is pieced together and satisfies the following formula:

[0011]

[0012]

[0013] in, Representative moment t KPI input, This represents the previous hidden state. This represents the backward hidden state at the next moment.

[0014] Furthermore, in the feature representation step, the semantic feature vector The [CLS] label vector output by the BERT pre-trained model satisfies the following formula:

[0015] in, s This represents the input log sequence.

[0016] Furthermore, in the feature representation step, the spatial feature vector It is obtained through the inter-layer propagation rules of graph convolutional networks, satisfying the following formula:

[0017] in, Representing the l The node feature matrix in a layered graph convolutional network. Representing the l The node feature matrix in a +1 layer graph convolutional network Represents a non-linear activation function. This represents the adjacency matrix of a self-loop. Degree matrix, Representing the l The learnable weight matrix of the layer.

[0018] Furthermore, in the knowledge fusion step, the knowledge-guided cross-attention calculation satisfies the following formula:

[0019]

[0020]

[0021]

[0022] in, Q Represents the query vector. K Represents the key vector. V Represents a value vector. , , These represent the weight matrices for the query vector, key vector, and value vector, respectively. Represents the dimension of the key vector.

[0023] Furthermore, in the knowledge fusion step, the modality confidence score Calculated using the following formula:

[0024] in, m Representing modal types, including KPI modality, log modality, and topology modality. This represents the feature vector of the corresponding mode. Represents the weight matrix. Represents the bias term. This represents the Sigmoid activation function.

[0025] Furthermore, in the knowledge fusion step, the multimodal fusion vector Calculated using the following formula:

[0026] in, Representative mode m The confidence score, Representative mode m The results of cross-attention Representative mode m eigenvectors, This represents a splicing operation.

[0027] Furthermore, in the fault prediction step, the total loss function Cross-entropy loss Physical consistency regularization term The composition satisfies the following formula:

[0028]

[0029] in, y Represents the true label, Represents the model's predicted value. Represents the physical loss weighting coefficient. R Representative domain rule set, Representative Rules r medium conditions A The probability of it being true. Representative fault B The probability of occurrence.

[0030] A fault prediction system for optical networks based on multimodal dynamic fusion includes: The data input module is used to synchronously collect key performance indicators of the optical network, namely KPI data stream, log text stream, and network topology stream, and to preprocess the three types of data streams. The feature representation module is used to extract the temporal feature vector of the KPI data stream through a bidirectional long short-term memory network, i.e., Bi-LSTM. The semantic feature vector of the log text stream is extracted using a BERT pre-trained model. The spatial feature vector of the network topology flow is extracted using a graph convolutional network, i.e., a graph network convolutional network (GCN). ; The knowledge fusion module is used to construct an optical network knowledge graph to obtain knowledge vectors. Based on the knowledge vector For the time-series feature vector semantic feature vector and spatial eigenvectors Knowledge-oriented cross-attention calculation is performed, and dynamic weighted fusion is combined with modality confidence scores to obtain a multimodal fusion vector. ; The fault prediction module is used to convert the multimodal fusion vector. The input is a multilayer perceptron (MLP) model, and the failure probability distribution is output through the Softmax function. At the same time, physical consistency loss is introduced to verify the prediction results.

[0031] The beneficial effects of this invention are as follows: (1) This invention adopts a multimodal dynamic fusion strategy based on a knowledge-oriented cross-attention mechanism. It actively searches for and aligns heterogeneous features using query vectors generated by optical network knowledge graphs, realizing a paradigm shift from passive feature splicing to active knowledge guidance. This design enables the model to deeply understand the fault evolution mechanism and significantly improves the accuracy and interpretability of fault diagnosis.

[0032] (2) This invention introduces a modal confidence assessment and dynamic weighting mechanism, which can quantify the information quality of each modal data stream in real time. When a modal data is missing or contaminated, the model can automatically reduce its weight and enhance the contribution of other effective modalities, giving the model strong elasticity, fault tolerance and robustness, and ensuring that it can still output stable prediction results in complex and ever-changing operation and maintenance environments.

[0033] (3) This invention innovatively introduces a physical consistency regularization term into the loss function, embedding the physical laws of optical communication as hard constraints into the model training and inference process. This not only forces the prediction results to conform to the physical common sense of optical network fault evolution and eliminates misjudgments that violate physical logic, but also plays a role in optimizing the solution space, effectively avoiding the model from getting trapped in local minima and accelerating the convergence process of the algorithm.

[0034] (4) This invention uses a parallel feature extraction architecture to independently model temporal, semantic and spatial features, thus fully preserving the essential characteristics of multi-source heterogeneous data in optical networks. Combined with a dynamic weight allocation strategy, the model can accurately capture potential fault features that are difficult to detect in a single mode, greatly improving the comprehensive judgment capability under multiple concurrent faults and complex scenarios.

[0035] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a diagram of the overall architecture of the MDF optical network fault prediction model. Detailed Implementation

[0037] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0038] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures, and should not be construed as limiting the invention. To better illustrate the embodiments of the present invention, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product; it is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0039] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0040] The present invention will be further described in detail below with reference to the accompanying drawings. This embodiment provides an optical network fault prediction method based on Multimodal Dynamic Fusion (MDF). This method is applied to data center optical network operation and maintenance scenarios, and its overall architecture is as follows: Figure 1 As shown, it includes a data input layer, a feature representation layer, a knowledge fusion layer, and a fault prediction layer.

[0041] Step 1: Data Input and Preprocessing The first step is to perform the data input process. Three types of heterogeneous data streams are collected synchronously through the optical network management system: Key Performance Indicator (KPI) data stream, log text stream, and network topology stream.

[0042] For the KPI data stream, which includes continuous time-series state quantities such as optical module transmit and receive power, operating voltage, bias current, and operating temperature, the preprocessing stage uses linear interpolation to repair missing values, and then zero-mean standardization is used to eliminate differences between different dimensions, achieving a normalized mapping of the original sequence.

[0043] For the log text stream, which includes system logs, proactive alarms, and operation records, a fixed-depth tree parsing algorithm is used for log template matching, automatically removing dynamic parameters such as timestamps, and combined with word segmentation cleaning and stop word filtering mechanisms to eliminate redundant and interfering information while retaining structured log semantics.

[0044] For network topology flow, the interconnection relationship and physical architecture of device nodes are detected in real time by relying on the link layer discovery protocol to construct a network topology diagram.

[0045] Step 2: Multimodal Feature Representation After data preprocessing is completed, the feature representation step is performed, using three parallel channels for independent modeling.

[0046] For KPI time-series feature extraction, a bidirectional long short-term memory (Bi-LSTM) network is employed. This network comprises two LSTM modules, forward and backward, capturing past and future contextual information respectively. For time series... t KPI input Forward hidden state With backward hidden state Updated using LSTM units respectively, as shown in Equation (1) and Equation (2): (1) (2) Final KPI feature vector The two are joined together as shown in formula (3): (3) For log semantic feature extraction, a BERT (Bidirectional Encoder Representations from Transformers) pre-trained model is used as the text representation engine. The pre-processed log sequence... s Input the BERT multi-layer Transformer encoder and extract the vector corresponding to the [CLS] label as the semantic feature vector of the entire log. As shown in formula (4): (4) For topological spatial feature extraction, network topology is modeled as a graph. A Graph Convolutional Network (GCN) is employed to capture the spatial propagation characteristics of faults. Let... To add a self-loop to the adjacency matrix, For degree matrix, For the first l The node feature matrix of the layer is given, and the inter-layer propagation rule is shown in formula (5): (5) in, Represents a non-linear activation function. Representing the l The learnable weight matrix of the layer, and the final output spatial feature vector. It integrates the node's own attributes and the fault status information of its multi-hop neighbors.

[0047] Step 3: Knowledge Graph Construction and Integration In the knowledge fusion step, a dedicated knowledge graph for optical networks is first constructed. The physical links and hierarchical dependencies of devices are extracted using a configuration management database, while fault causal relationships and alarm correlations are extracted using an operations and maintenance expert knowledge base. Triples are then learned using the TransE algorithm. H , R , T The low-dimensional vector representation of entities and relations in the model is used to minimize the objective function using a loss function based on interval sorting, as shown in Equation (10): (10) in, Denotes the set of positive sample triples. Denotes the set of negative sample triples. d For distance measurement function, The boundary hyperparameters are used to dynamically update and generate knowledge vectors. .

[0048] Subsequently, a knowledge-oriented cross-attention mechanism is used to achieve dynamic fusion. This involves integrating knowledge vectors... The mapping is to the query vector Q, as shown in formula (6): (6) The multimodal data features are concatenated and mapped to a key vector K and a value vector. V As shown in formulas (7) and (8): (7) V (8) in , , These represent the weight matrices. The attention weights are calculated by scaling the dot product attention, as shown in formula (9): (9) in Represents the dimension of the key vector.

[0049] Simultaneously, a modal confidence evaluation mechanism is introduced, targeting three modalities: KPI, logs, and topology. m Calculate their respective confidence scores using the Sigmoid activation function. As shown in formula (11): (11) in Represents the weight matrix. This represents the bias term.

[0050] Final multimodal fusion vector The cross-attention results are dynamically weighted and fused based on the confidence scores, as shown in formula (12): (12) This mechanism ensures that when data for one modality is missing or of poor quality, the model can be effectively compensated for by other modalities.

[0051] Step 4: Fault Prediction and Physical Verification Finally, the fault prediction step is performed. The fused vector... The input is processed by a multilayer perceptron (MLP) and subjected to a high-order nonlinear transformation. The probability distribution of various faults is then output through a softmax function. y As shown in formula (13): (13) in The weight matrix represents the weights of the fully connected layer. b This represents the bias vector of the fully connected layer.

[0052] To ensure the physical validity of the prediction results, the traditional cross-entropy loss... Based on this, add a physical consistency regularization term. For the set of domain rules R When the rules r Defined conditions A When established, a fault should necessarily be identified. B The occurrence of this regularization term is defined as shown in formula (14): (14) Total loss function The weighted sum of the above two items is shown in formula (15): (15) in This is the physical loss weighting coefficient. y Represents the true label, This represents the model's predicted value. The physical consistency loss, acting as a strong convex regularization constraint, reduces the solution space and accelerates the convergence speed of the gradient descent algorithm, ensuring that the model output is both accurate and conforms to the physical laws of optical communication.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A fault prediction method for optical networks based on multimodal dynamic fusion, characterized in that: Includes the following steps: Data input steps: Synchronously collect key performance indicators of the optical network, namely KPI data stream, log text stream, and network topology stream, and preprocess the three types of data streams; Feature representation steps: Extract the temporal feature vector of the KPI data stream using a bidirectional long short-term memory network, i.e., Bi-LSTM. The semantic feature vector of the log text stream is extracted using a BERT pre-trained model. The spatial feature vector of the network topology flow is extracted using a graph convolutional network, i.e., a graph convolutional network (GCN). ; Knowledge fusion steps: Construct an optical network knowledge graph to obtain knowledge vectors Based on the knowledge vector For the time-series feature vector semantic feature vector and spatial eigenvectors Knowledge-oriented cross-attention calculation is performed, and dynamic weighted fusion is combined with modality confidence scores to obtain a multimodal fusion vector. ; Fault prediction step: The multimodal fusion vector The input is a multilayer perceptron (MLP) model, and the failure probability distribution is output through the Softmax function. At the same time, physical consistency loss is introduced to verify the prediction results.

2. The optical network fault prediction method based on multimodal dynamic fusion according to claim 1, characterized in that: The data input step specifically includes: The KPI data stream is repaired by linear interpolation and then normalized to zero mean. The log text stream is subjected to a fixed-depth tree parsing algorithm for log template matching, followed by word segmentation, cleaning, and stop word filtering. The network topology flow relies on the link layer discovery protocol to detect the interconnection relationship and physical architecture of device nodes in real time.

3. The optical network fault prediction method based on multimodal dynamic fusion according to claim 1, characterized in that: In the feature representation step, the temporal feature vector From forward hidden state and backward hidden state It is pieced together and satisfies the following formula: in, Representative moment t KPI input, This represents the previous hidden state. This represents the backward hidden state at the next moment.

4. The optical network fault prediction method based on multimodal dynamic fusion according to claim 1, characterized in that: In the feature representation step, the semantic feature vector The [CLS] label vector output by the BERT pre-trained model satisfies the following formula: in, s This represents the input log sequence.

5. The optical network fault prediction method based on multimodal dynamic fusion according to claim 1, characterized in that: In the feature representation step, the spatial feature vector It is obtained through the inter-layer propagation rules of graph convolutional networks, satisfying the following formula: in, Representing the l The node feature matrix in a layered graph convolutional network, Representing the l The node feature matrix in a +1 layer graph convolutional network Represents a non-linear activation function. This represents the adjacency matrix of a self-loop. Degree matrix, Representing the l The learnable weight matrix of the layer.

6. The optical network fault prediction method based on multimodal dynamic fusion according to claim 1, characterized in that: In the knowledge fusion step, the knowledge-oriented cross-attention calculation satisfies the following formula: in, Q Represents the query vector. K Represents the key vector. V Represents a value vector. , , These represent the weight matrices for the query vector, key vector, and value vector, respectively. Represents the dimension of the key vector.

7. The optical network fault prediction method based on multimodal dynamic fusion according to claim 1, characterized in that: In the knowledge fusion step, the modality confidence score Calculated using the following formula: in, m Representing modal types, including KPI modality, log modality, and topology modality. This represents the feature vector of the corresponding mode. Represents the weight matrix. Represents the bias term. This represents the Sigmoid activation function.

8. The optical network fault prediction method based on multimodal dynamic fusion according to claim 1, characterized in that: In the knowledge fusion step, the multimodal fusion vector Calculated using the following formula: in, Representative mode m The confidence score, Representative mode m The results of cross-attention Representative mode m eigenvectors, This represents a splicing operation.

9. The optical network fault prediction method based on multimodal dynamic fusion according to claim 1, characterized in that: In the fault prediction step, the total loss function Cross-entropy loss Physical consistency regularization term The composition satisfies the following formula: in, y Represents the true label, Represents the model's predicted value. Represents the physical loss weighting coefficient. R Representative domain rule set, Representative Rules r medium conditions A The probability of it being true. Representative fault B The probability of occurrence.

10. A fault prediction system for optical networks based on multimodal dynamic fusion, characterized in that: include: The data input module is used to synchronously collect key performance indicators of the optical network, namely KPI data stream, log text stream, and network topology stream, and to preprocess the three types of data streams. The feature representation module is used to extract the temporal feature vector of the KPI data stream through a bidirectional long short-term memory network, i.e., Bi-LSTM. The semantic feature vector of the log text stream is extracted using a BERT pre-trained model. The spatial feature vector of the network topology flow is extracted using a graph convolutional network, i.e., a graph convolutional network (GCN). ; The knowledge fusion module is used to construct an optical network knowledge graph to obtain knowledge vectors. Based on the knowledge vector For the time-series feature vector semantic feature vector and spatial eigenvectors Knowledge-oriented cross-attention calculation is performed, and dynamic weighted fusion is combined with modality confidence scores to obtain a multimodal fusion vector. ; The fault prediction module is used to convert the multimodal fusion vector. The input is a multilayer perceptron (MLP) model, and the failure probability distribution is output through the Softmax function. At the same time, physical consistency loss is introduced to verify the prediction results.

Images