A fgOTN communication network fault prediction method based on topology calculation

Abstract

The application relates to the technical field of communication fault prediction, in particular to a fgOTN communication network fault prediction method based on topology calculation, which comprises the following steps: collecting real-time performance data, physical connection relations and OSU configurations of the fgOTN communication network, and constructing a physical layer matrix and a service layer matrix; generating a fault risk conduction factor based on the time division multiplexing frame structure of the fgOTN, and constructing an interlayer mapping correlation matrix; combining the above matrix to establish a joint topology matrix, and executing Laplace decomposition to generate a topology position identifier; constructing a space-time state diagram based on the topology position identifier, executing feature calculation containing vertical aggregation and horizontal diffusion by using a graph convolution operator, and generating a node feature vector; constructing a space-time state sequence based on a time sliding window mechanism and inputting the space-time state sequence into a time slot analysis network, and generating a fault prediction result. By establishing the quantitative mapping of the physical layer and the service layer, the application realizes the unified modeling of cross-layer fault conduction and same-layer risk diffusion.

Description

Technical Field

[0001] This invention relates to the field of communication fault prediction technology, specifically to a fault prediction method for fgOTN communication networks based on topology calculation. Background Technology

[0002] As communication networks evolve towards all-optical networks, fgOTN technology, with its hard isolation and flexible scheduling capabilities for small-granularity services, has become the preferred solution in key areas such as power and 5G network infrastructure. In the fgOTN network architecture, the physical layer optical path and the service layer logical channel are tightly coupled through a complex time-division multiplexing mechanism. A single physical port often carries multiple OSU services with different bandwidth requirements, and ensuring their high-reliability transmission is the core requirement for network operation and maintenance. Currently, fault prediction technologies for communication networks mainly rely on statistical methods or deep learning algorithms to predict future states by analyzing historical performance data. These methods typically treat network devices as independent nodes or perform time-series analysis only on data from a single layer, lacking a cross-layer fault propagation characterization mechanism based on fine-grained time-slot resources. Specifically, the physical layer and the service layer of fgOTN are not simply linearly corresponding, but rather have a dynamic mapping relationship based on time slots. The impact of minor fluctuations in physical layer performance on upper-layer services depends strictly on the dynamic time slot position and number occupied by the service in the physical frame structure. Existing prediction methods often treat the physical topology and the service logic topology separately. This lack of cross-layer correlation features makes it impossible for the model to accurately capture the transmission path and evolution trend of faults when faced with potential service layer problems caused by slight degradation of the physical layer, which seriously affects the accuracy and timeliness of fault prediction.

[0003] To address this, a fault prediction method for fgOTN communication networks based on topology calculation is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a fault prediction method for fgOTN communication networks based on topology calculation. By establishing a quantitative mapping between the physical layer and the service layer, a unified modeling of cross-layer fault propagation and intra-layer risk diffusion is achieved.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Collect real-time performance data, physical connection relationships, and OSU configuration of the fgOTN communication network to construct the physical layer matrix and service layer matrix; The dynamic time slot occupancy rate of OSU services on physical ports is calculated based on the time-division multiplexing frame structure of fgOTN and converted into a fault risk transmission factor. The fault risk transmission factor is used as an element value to construct an inter-layer mapping correlation matrix. A joint topology matrix is ​​established by combining the physical layer matrix and the service layer matrix. Laplace decomposition is performed on the joint topology matrix to generate topology location identifiers. A spatiotemporal state graph is constructed based on topological location identifiers; based on the spatiotemporal state graph, graph convolution operators are used to perform topological feature calculations to generate node feature vectors; the topological feature calculations include vertical and horizontal dimensions, the vertical dimension projects the inter-layer mapping correlation matrix to the OSU service layer by weighting the fault risk transmission factor, and the horizontal dimension defines the topological receptive field based on the service layer matrix and aggregates the neighborhood structure distribution features. The node feature vectors generated by collecting multiple consecutive time slices based on the time sliding window mechanism are used to construct a spatiotemporal state sequence and input into the time slot analysis network to generate the fgOTN communication network fault prediction result.

[0006] Preferably, the physical layer matrix is ​​a physical adjacency matrix based on the physical connection relationships of the fgOTN communication network, and its dimension is [missing information]. ,in The matrix element represents the number of physical nodes. The elements are determined based on the physical connection relationship. The element value is 1 when there is a physical link between two physical nodes, and 0 otherwise. The business layer matrix is ​​a logical adjacency matrix representing the logical connection relationships of OSU business operations, with dimensions of [missing information]. ,in The number of service nodes is determined by the matrix elements based on the OSU configuration. The element value is 1 when there is an OSU logical channel between two service nodes, and 0 otherwise.

[0007] Preferably, the specific process of calculating the dynamic time slot occupancy rate of OSU services on physical ports and converting it into a fault risk transmission factor based on the fgOTN-based time slot multiplexing mechanism is to construct an index mapping table between physical ports and OSU services based on the OSU configuration. Based on the time-division multiplexing frame structure of fgOTN, the time slot allocation bitmap of the physical port in a single multiplexing cycle is parsed. Based on the time slot allocation bitmap, the actual location and number of tributary time slots occupied by each OSU service carried on the physical port are identified. The ratio of the number of tributary time slots to the total time slot capacity of the physical port frame structure payload area is calculated to generate a dynamic time slot occupancy rate. The dynamic time slot occupancy rate is mapped to a fault risk transmission factor through a nonlinear weighting function.

[0008] Preferably, the inter-layer mapping correlation matrix has a dimension of The heterogeneous correlation matrix, where The number of physical nodes. This refers to the number of business nodes. The element values ​​in the inter-layer mapping association matrix are the fault risk transmission factors for physical nodes and service nodes with a bearing relationship, located at the corresponding intersection positions in the inter-layer mapping association matrix; for physical nodes and service nodes without a bearing relationship, the element values ​​at the corresponding intersection positions in the inter-layer mapping association matrix are 0. A joint topology matrix is ​​constructed using a block matrix structure. The physical layer matrix is ​​located in the upper left diagonal block of the joint topology matrix, and the service layer matrix is ​​located in the lower right diagonal block of the joint topology matrix. The inter-layer mapping association matrix is ​​located in the upper right off-diagonal block of the joint topology matrix, and the transpose of the inter-layer mapping association matrix is ​​located in the lower left off-diagonal block of the joint topology matrix.

[0009] Preferably, the Laplace decomposition is performed by determining the connection density of network nodes based on the joint topology matrix, generating connection metric parameters that reflect the hub status of nodes in the network, and using the connection metric parameters to standardize the network connection relationships and establish an association model that reflects the global connection architecture of the fgOTN network. Structural features are extracted from the association model, the connection patterns and link coupling relationships of network nodes are analyzed, and topological location identifiers that identify the distribution characteristics of each physical node and service node in the network topology are generated according to the order of their impact on network connectivity.

[0010] Preferably, the specific process of constructing the spatiotemporal state graph based on topological location identifiers is as follows: using a joint topological matrix as the edge set of the graph, and using physical nodes and service nodes as the node set of the graph, a topological skeleton of the graph is established; the topological location identifiers are concatenated with the real-time performance data by feature dimensions to generate node attribute vectors; the real-time performance data includes input optical power, output optical power, optical signal-to-noise ratio and bit error rate for physical nodes, and OSU channel latency, latency jitter and OSU frame loss rate for service nodes; The node attribute vectors are assigned to the corresponding nodes in the node set of the graph to construct a spatiotemporal state graph.

[0011] Preferably, the use of graph convolution operators to perform topology-based feature computation specifically includes vertical aggregation and horizontal diffusion; The vertical aggregation is achieved by using the graph convolution operator with the inter-layer mapping correlation matrix as the adjacency weight matrix for cross-layer convolution, and projecting the state features of the physical layer nodes to the corresponding OSU service layer nodes according to the fault risk transmission factor; the horizontal diffusion is achieved by using the graph convolution operator to determine the convolution receptive field based on the topological connection relationship defined by the service layer matrix, and performing neighborhood feature aggregation on the OSU service layer nodes.

[0012] Preferably, the time slot analysis network specifically associates the extracted spatiotemporal state sequence with the transmission time slots of fgOTN, injects a time sequence identifier corresponding to the transmission period of the OSU service frame, and characterizes the positional relationship of different transmission time slots in the multiframe sequence. Based on the traffic fluctuation correlation analysis mechanism, the performance parameter change pattern of different time slots in the historical transmission cycle is detected, the time sequence anomaly pattern is identified, and the time slot significance weights that reflect the coupling relationship of traffic characteristics between time slots are generated. The spatiotemporal state sequence is weighted according to the transmission time slot using the time slot significance weight, and a time slot-aware context vector reflecting the evolution trend of network channel state is extracted. The fault state of the current fgOTN communication link is determined based on the time slot-aware context vector.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention analyzes the time-division multiplexing frame structure of fgOTN, calculates the dynamic time slot occupancy rate of OSU services on physical ports, and converts it into a fault risk transmission factor to construct an inter-layer mapping correlation matrix. This mechanism breaks through the limitations of existing technologies that can only perform simple physical-logical correspondence or binary correlation. By introducing the time slot occupancy rate as a key intermediate variable, it establishes a quantitative bridge between the physical layer and the service layer. It can dynamically reflect the degree of nonlinear impact of physical layer performance on the differences in different services based on the actual resource weights occupied by the services, thereby characterizing the transmission characteristics of fault risks in the cross-layer architecture and providing a computational basis for the model to perceive the potential risks of upper-layer services caused by physical layer degradation.

[0014] 2. This invention constructs a joint topology matrix comprising a physical layer matrix, a service layer matrix, and an inter-layer mapping correlation matrix, and utilizes graph convolution operators to perform feature calculations in both vertical and horizontal dimensions. This design unifies the previously fragmented physical network architecture and OSU service logic architecture into a single high-dimensional topology space for modeling. Specifically, vertical dimension calculation simulates the cross-layer penetration of fault risks from physical facilities to service channels through weighted projection, while horizontal dimension calculation simulates the diffusion effect of risks within the same layer's network neighborhood. This comprehensive spatial feature extraction method enables the prediction model to not only identify single-point faults but also effectively capture network-level structural vulnerabilities, solving the problem of inconsistent fault feature representation under heterogeneous topologies.

[0015] 3. This invention constructs a time-slot analysis network, associates the extracted spatiotemporal state sequence with the fgOTN transmission time slots, and generates time slot significance weights using a traffic fluctuation correlation analysis mechanism. Since fgOTN is based on time-slot scheduling, faults often exhibit regular abnormal patterns within specific time-slot periods. This invention utilizes time sequence identifiers and significance weights to analyze the changing patterns of performance parameters at the microsecond-level time-slot level, rather than merely remaining at the macroscopic time series statistics. This enables the model to identify temporal anomaly patterns hidden in complex multiframe sequences, thereby achieving the identification and prediction of fault trends that gradually accumulate and evolve over time on specific time-slot resources. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a fault prediction method for fgOTN communication networks based on topology calculation. Figure 2 This is a schematic diagram of the fault risk factor calculation process based on time slot bitmap determination; Figure 3 This is a schematic diagram of the vertical assembly logic flow for the physical-business joint topology matrix. Detailed Implementation

[0017] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: Please see Figure 1 This invention provides a fault prediction method for fgOTN communication networks based on topology calculation, and the technical solution is as follows: A fault prediction method for fgOTN communication networks based on topology calculation includes: Collect real-time performance data, physical connection relationships, and OSU configuration of the fgOTN communication network to construct the physical layer matrix and service layer matrix; The dynamic time slot occupancy rate of OSU services on physical ports is calculated based on the time-division multiplexing frame structure of fgOTN and converted into a fault risk transmission factor. The fault risk transmission factor is used as an element value to construct an inter-layer mapping correlation matrix. A joint topology matrix is ​​established by combining the physical layer matrix and the service layer matrix. Laplace decomposition is performed on the joint topology matrix to generate topology location identifiers. A spatiotemporal state graph is constructed based on topological location identifiers; based on the spatiotemporal state graph, graph convolution operators are used to perform topological feature calculations to generate node feature vectors; the topological feature calculations include vertical and horizontal dimensions, the vertical dimension projects the inter-layer mapping correlation matrix to the OSU service layer by weighting the fault risk transmission factor, and the horizontal dimension defines the topological receptive field based on the service layer matrix and aggregates the neighborhood structure distribution features. The node feature vectors generated by collecting multiple consecutive time slices based on the time sliding window mechanism are used to construct a spatiotemporal state sequence and input into the time slot analysis network to generate the fgOTN communication network fault prediction result.

[0019] Specifically, the process of collecting real-time performance data, physical connection relationships, and OSU configurations of the fgOTN communication network, and constructing the physical layer matrix and service layer matrix, involves performing a full network resource scan of the fgOTN communication network, identifying all physical nodes in the network and service nodes participating in service scheduling, and establishing a node index system. Establishing the node index system involves counting the total number of physical nodes in the entire network and assigning a unique numerical index to each physical node; simultaneously, counting the total number of service nodes in the entire network and assigning a unique numerical index to each service node. Establish a real-time monitoring and acquisition channel to obtain performance index data of the physical layer and the service layer; for each physical node, collect its real-time input optical power, output optical power, optical signal-to-noise ratio and bit error rate data by reading the device port register or performance management database; for each service node, collect the channel latency, latency jitter and OSU frame loss rate data of the OSU channel through in-band measurement or operation and maintenance frame monitoring mechanism, and use these data as the basis for subsequent node attribute assignment. The physical layer matrix is ​​a physical adjacency matrix based on the physical connection relationships of the fgOTN communication network, and its dimension is [missing information]. ,in The matrix element represents the number of physical nodes. The elements are determined based on the physical connection relationship. The element value is 1 when there is a physical link between two physical nodes, and 0 otherwise. When constructing the physical layer matrix, each row and column of the matrix are traversed. Based on the network topology data, it is determined whether there is a physical link between the physical node corresponding to the row index and the physical node corresponding to the column index: when there is a direct physical link between the two physical nodes, the element value of the physical layer matrix position is set to 1; otherwise, when there is no physical link between the two physical nodes, the element value of the physical layer matrix position is set to 0. The business layer matrix is ​​a logical adjacency matrix representing the logical connection relationships of OSU business operations, with dimensions of [missing information]. ,in The number of service nodes is determined by the matrix elements based on the OSU configuration. The element value is 1 when there is an OSU logical channel between two service nodes, and 0 otherwise. When constructing the business layer matrix, it traverses each row and column of the matrix, parses the OSU configuration data, and determines whether there is a configured OSU logical channel between the business node corresponding to the row index and the business node corresponding to the column index: when there is an OSU logical channel between the two business nodes, the element value at the position of the business layer matrix is ​​set to 1; otherwise, when there is no OSU logical channel between the two business nodes, the element value at the position of the business layer matrix is ​​set to 0. By establishing a network-wide resource scanning and node indexing system, this invention achieves comprehensive digital identification of physical entities and logical services in the fgOTN network, ensuring the integrity of the monitoring scope. Furthermore, by constructing independent physical layer and service layer matrices, the complex network connection relationships are transformed into a structured mathematical model that can be directly processed by a computer. This two-layer matrix structure not only clearly decouples and characterizes the topological features of the underlying optical transmission links and the upper-layer OSU service channels, but also, combined with real-time collected multi-dimensional performance indicators, provides a data foundation for subsequent cross-layer topology calculation and fault feature mining based on graph neural networks. The specific process of calculating the dynamic time slot occupancy rate of OSU services on physical ports and converting it into a fault risk transmission factor based on the fgOTN-based time slot multiplexing mechanism is to construct an index mapping table between physical ports and OSU services based on the OSU configuration. See Figure 2 Based on the time-division multiplexing frame structure of fgOTN, the time slot allocation bitmap of the physical port in a single multiplexing cycle is parsed. Based on the time slot allocation bitmap, the actual location and number of tributary time slots occupied by each OSU service carried on the physical port are identified. The ratio of the number of tributary time slots to the total time slot capacity of the physical port frame structure payload area is calculated to generate a dynamic time slot occupancy rate. The dynamic time slot occupancy rate is mapped to a fault risk transmission factor through a nonlinear weighting function.

[0020] Specifically, the process of constructing an index mapping table between physical ports and optical service units based on optical service unit configuration involves first traversing each physical node and its associated physical port in the fine-grained optical transport network and reading the service routing table; for each physical port, retrieving all logical channel identifiers of optical service units flowing through that port and establishing a one-to-many hierarchical index relationship; the index mapping table uses the globally unique identifier of the physical port as the primary key and the list of identifiers of all optical service units currently carried by the physical port as the data value, thus clarifying the subordinate bearing relationship between physical layer resources and service layer logical channels; The process of parsing the time slot allocation bitmap of the physical port in a single multiplexing cycle based on the time-division multiplexing frame structure of the fine-grained optical transport network involves reading the overhead region of the optical transport network frame structure received by the physical port; specifically, locating the multiplexing structure identifier or payload structure identifier region in the frame structure and extracting a binary bitmap sequence used to indicate the time slot allocation status of the payload region; the length of the binary bitmap sequence corresponds to the total time slot capacity of the physical port in a single multiplexing cycle, each binary bit in the bitmap corresponds to a basic time slot unit, and its numerical status is used to indicate whether the time slot is currently occupied by a specific optical service unit or is in an idle state. The process of identifying the actual location and number of tributary time slots occupied by each optical service unit carried on the physical port is based on the index mapping table established above. For each optical service unit in the mapping table, pattern matching and counting are performed in the time slot allocation bitmap. Specifically, the channel number identifier in the bitmap is parsed, and the total number of binary bits exclusively occupied by a specific optical service unit in the current multiplexing cycle is counted, which is used as the number of tributary time slots occupied by the optical service unit. At the same time, the specific sequence number position of these time slots in the frame structure payload area is recorded to determine the time distribution characteristics of the service on the physical channel. The process of calculating the ratio of the number of branch time slots to the total time slot capacity of the physical port frame structure payload area and generating the dynamic time slot occupancy rate uses the number of branch time slots occupied by the optical service unit obtained above as the numerator and the maximum total time slot capacity that the physical port frame structure payload area can provide as the denominator. A division operation is performed to obtain a value between zero and one, which is the dynamic time slot occupancy rate. This quantifies the degree of consumption of physical port transmission resources by a single optical service unit and reflects the bandwidth ratio of the service in the physical channel. The process of mapping the dynamic time slot occupancy rate to the fault risk transmission factor through a nonlinear weighting function is to select the Sigmoid function as the mapping algorithm and directly use the currently calculated dynamic time slot occupancy rate as the input variable of the Sigmoid function to generate the fault risk transmission factor. Its value increases with the increase of the dynamic time slot occupancy rate. In this embodiment, a sigmoid function with an amplification factor of 4 is specifically selected as the mapping algorithm. The calculation process involves subtracting 0.5 from the dynamic time slot occupancy rate, multiplying by -4, calculating its natural exponent, adding 1, and finally taking the reciprocal of the result. This parameter setting ensures that when the time slot occupancy rate exceeds 0.5, the output fault risk transmission factor exhibits a steep upward trend, thereby sensitively quantifying the sensitivity of service bandwidth usage to physical layer performance fluctuations within the numerical range of 0 to 1. By analyzing the time slot bitmap, the physical resource occupancy of optical service units is accurately calculated, achieving precise mapping at the micro-time slot level. Furthermore, by utilizing the normalization and monotonically increasing characteristics of the Sigmoid function, the time slot occupancy rate is transformed into a standardized fault risk transmission factor, objectively quantifying the nonlinear law that higher resource consumption leads to greater risk, and providing characteristic data for cross-layer fault prediction.

[0021] See Figure 3 The fault risk transmission factor is used as an element value to construct an inter-layer mapping correlation matrix, and the inter-layer mapping correlation matrix has a dimension of The heterogeneous correlation matrix, where The number of physical nodes. This refers to the number of business nodes. The element values ​​in the inter-layer mapping association matrix are the fault risk transmission factors for physical nodes and service nodes with a bearing relationship, located at the corresponding intersection positions in the inter-layer mapping association matrix; for physical nodes and service nodes without a bearing relationship, the element values ​​at the corresponding intersection positions in the inter-layer mapping association matrix are 0. A joint topology matrix is ​​constructed using a block matrix structure. The physical layer matrix is ​​located in the upper left diagonal block of the joint topology matrix, and the service layer matrix is ​​located in the lower right diagonal block of the joint topology matrix. The inter-layer mapping association matrix is ​​located in the upper right off-diagonal block of the joint topology matrix, and the transpose of the inter-layer mapping association matrix is ​​located in the lower left off-diagonal block of the joint topology matrix.

[0022] Specifically, create a heterogeneous association matrix with dimensions N × M, where N represents the total number of physical nodes and M represents the total number of business nodes; initialize all elements of this matrix to the value 0. The process of filling the element values ​​in the inter-layer mapping association matrix involves traversing each intersection position of the matrix and performing logical judgments and assignments; specifically, for the first element in the matrix... Line 1 The elements of the column (where) The value range is [1, N]. The value range is [1, M]). The index mapping table between physical ports and optical service units is consulted to determine whether the physical node represented by the row index is the actual bearer of the service node represented by the column index. If it is determined that there is a bearer relationship between the two, the fault risk transmission factor of the service to the physical port calculated in the previous steps is called, and the fault risk transmission factor is filled into the corresponding row and column intersection position. If it is determined that there is no bearer relationship between the two, the element value at this position is kept to be 0. The process of constructing a joint topology matrix using a block matrix structure aims to unify and integrate the physical layer, business layer, and inter-layer relationships into a single topology space; specifically, it defines a dimension as... The global joint matrix is ​​divided into four logical quadrants: the upper left region is designated as the physical domain block, the lower right region as the business domain block, the upper right region as the uplink mapping block, and the lower left region as the downlink feedback block. The specific operation of filling each block of the joint topology matrix is ​​as follows: The physical layer matrix is ​​completely embedded into the upper left diagonal block to represent the connection topology within the physical layer; the service layer matrix is ​​completely embedded into the lower right diagonal block to represent the logical topology within the service layer; and the dimension constructed in this embodiment is... The inter-layer mapping correlation matrix is ​​embedded into the upper right off-diagonal block; The process of processing the lower left off-diagonal block of the joint topology matrix involves performing a transpose operation on the inter-layer mapping correlation matrix; specifically, transposing the matrix with dimension 1... The rows and columns of the interlayer mapping correlation matrix are interchanged to generate a dimension of . The transpose of the matrix is ​​then embedded into the lower left off-diagonal block of the joint topological matrix, giving the joint topological matrix symmetry about the main diagonal. By constructing an inter-layer mapping association matrix, the heterogeneous mapping problem of inconsistent node scale between physical and business networks is effectively solved. Furthermore, by using a block structure to construct a joint topology matrix, the two-layer network is integrated into a unified graph structure. This not only opens up the digital path for physical faults to propagate to the business layer, but also maintains the mathematical symmetry of the topology by embedding a transpose matrix, enabling the graph algorithm to simultaneously calculate intra-layer diffusion and inter-layer penetration within a closed-loop space.

[0023] The Laplace decomposition is performed to determine the connection density of network nodes based on the joint topology matrix, generate connection metric parameters that reflect the hub status of nodes in the network, and use the connection metric parameters to standardize the network connection relationships to establish an association model that reflects the global connection architecture of the fgOTN network. Structural features are extracted from the association model, the connection patterns and link coupling relationships of network nodes are analyzed, and topological location identifiers that identify the distribution characteristics of each physical node and service node in the network topology are generated according to the order of their impact on network connectivity.

[0024] The process of determining the connection density of network nodes based on the joint topology matrix and generating connection metric parameters reflecting the hub status of nodes in the network involves performing a row-wise summation operation on the joint topology matrix; traversing each row of the matrix and accumulating the element values ​​of all columns in that row, the sum being the degree value of the corresponding node, quantifying the connection breadth and weight of the node in the entire fgOTN network; constructing a diagonal matrix with the same dimension as the joint topology matrix, and sequentially filling the main diagonal positions of the calculated degree value of each node into the diagonal matrix, filling the remaining positions with the value 0, thus generating a degree matrix; transforming the complex network topology relationship into quantifiable connection metric parameters, realizing a numerical representation of the comprehensive hub status of nodes in the physical-service two-layer architecture.

[0025] The inverse square root operation is performed on the main diagonal elements of the degree matrix, i.e., the square root of the reciprocal of each degree value is taken to generate the inverse square root matrix of the degree matrix. Next, the inverse square root matrix is ​​used to perform bilateral matrix multiplication on the joint topology matrix, i.e., multiplying the joint topology matrix by the inverse square root matrix on both the left and right sides respectively to obtain the normalized adjacency matrix. Finally, an identity matrix of the same dimension is constructed, and the normalized adjacency matrix is ​​subtracted from the identity matrix. The result is the normalized Laplace matrix. The Laplace matrix eliminates the scale effect of node degree differences on the graph structure, and constitutes an association model that accurately reflects the global connectivity architecture of the network. Matrix decomposition is performed on the standardized Laplacian matrix to directly extract a set of eigenvalues ​​of the matrix and the eigenvector corresponding to each eigenvalue; wherein, the magnitude of the eigenvalue represents the importance of the connection pattern represented by the corresponding eigenvector in the network topology, and the eigenvector describes the spatial distribution of network nodes in the connection pattern. All acquired feature values ​​are sorted in ascending order. In this embodiment, the 2nd to K+1th non-zero feature values ​​with the smallest values ​​after sorting are selected (K is the preset embedding dimension, which is 16 in this embodiment, and this value is set to match the memory alignment specification and tensor computation parallelism of the underlying hardware). K feature vectors corresponding to these K feature values ​​are extracted, and these feature vectors are combined column by column to form a new feature matrix. Each row of data is the topological location identifier of the corresponding physical node or business node. Through this spectral clustering mapping method, the topological location identifier vectors of nodes that are closely connected or have similar structures in the network topology are closer in Euclidean space. Considering that the smallest eigenvalue of the standardized Laplacian matrix is ​​always 0 and cannot distinguish node differences, this embodiment explicitly discards the first eigenvalue after sorting, and instead selects the eigenvectors corresponding to the second to seventeenth smallest non-zero eigenvalues. This method constructs a 16-dimensional topological location identifier, suitable for a network size of approximately 50 nodes. It effectively captures global topological connectivity features while avoiding computational noise introduced by excessively high dimensionality. By utilizing standardized Laplacian matrix decomposition and based on the principle of spectral clustering, high-dimensional sparse joint topology is mapped to low-dimensional dense topological location identifiers. This effectively eliminates the influence of node degree differences, uncovers deep connection patterns, and ensures that nodes that are densely connected in the topology remain adjacent in the feature space.

[0026] The specific process of constructing a spatiotemporal state graph based on topological location identifiers involves establishing a topological skeleton of the graph by using a joint topological matrix as the edge set and physical nodes and service nodes as the node set; concatenating the topological location identifiers with the real-time performance data by feature dimensions to generate node attribute vectors; the real-time performance data includes input optical power, output optical power, optical signal-to-noise ratio, and bit error rate for physical nodes, and OSU channel latency, latency jitter, and OSU frame loss rate for service nodes. The node attribute vectors are assigned to the corresponding nodes in the node set of the graph to construct a spatiotemporal state graph.

[0027] Specifically, instantiate a graph data structure object; define a node list for the graph, including all physical nodes and all business nodes, with the total number of nodes equal to the sum of the number of physical nodes and the number of business nodes; load the joint topology matrix as an adjacency matrix, traverse each element in the matrix, and when the value of the element at the intersection of a row and a column is not zero, create a directed edge or weighted edge between the corresponding two nodes in the graph structure, and use the matrix element value as the weight attribute of the edge; The process of concatenating topology location identifiers with real-time performance data along feature dimensions to generate node attribute vectors involves first preprocessing and normalizing the real-time performance data. For physical nodes, four indicators are read: input optical power, output optical power, optical signal-to-noise ratio, and bit error rate. Since the dimensions of each indicator are different, maximum-minimum normalization is used to map them to a unified numerical range. For service nodes, three indicators are read: optical service unit channel latency, latency jitter, and frame loss rate, and normalization is also performed. To solve the problem of heterogeneous node attribute alignment, this embodiment constructs a performance feature container of uniform length. For physical nodes, the four indicators are filled into the beginning of the container, and zeros are padded to the end. For service nodes, the three indicators are filled into the end of the container, and zeros are padded to the beginning. Perform feature concatenation operation. For each node in the graph, extract the topological location identifier vector corresponding to the node generated by Laplacian decomposition. Concatenate the real-time performance feature container processed above directly to the end of the topological location identifier vector to form a high-dimensional combined vector. The first half encodes the node's global structural role and hub status in the whole network, and the second half encodes the node's dynamic operating quality at the current moment. The two together constitute the node attribute vector of the node. Traverse each node object in the graph structure and write the node attribute vector generated above into the feature attribute field of the node object; at this time, the graph data structure not only contains the connection relationship defined by the joint topology matrix, but also contains the node state information defined by the node attribute vector; the final spatiotemporal state graph is a tensor object containing complete topological information and feature information, and is tagged with the timestamp of the current acquisition time. By constructing a spatiotemporal state diagram, static topology features and dynamic performance features are deeply integrated at the node level. Among them, the topology location identifier provides the spatial coordinate information of the node in the global network, while the real-time performance data provides the current running status information of the node. This splicing method enables the subsequent prediction model to not only perceive what happened to a node, but also to perceive where a node is.

[0028] The use of graph convolution operators to perform topology-based feature computation specifically includes vertical aggregation and horizontal diffusion. The vertical aggregation is achieved by using the graph convolution operator with the inter-layer mapping correlation matrix as the adjacency weight matrix for cross-layer convolution, and projecting the node feature vectors of the physical layer nodes to the corresponding OSU service layer nodes according to the fault risk transmission factor; the horizontal diffusion is achieved by using the graph convolution operator to determine the convolution receptive field based on the topological connection relationship defined by the service layer matrix, and performing neighborhood feature aggregation on the OSU service layer nodes.

[0029] Specifically, the process of performing topology-based feature calculation using graph convolution operators involves constructing a multi-channel graph neural network layer in the computation model, designed as a parallel processing mechanism, including a vertical aggregation channel specifically for handling cross-layer influences and a horizontal diffusion channel specifically for handling same-layer propagation. The vertical aggregation is a process that uses graph convolution operators with the inter-layer mapping correlation matrix as the adjacency weight matrix for cross-layer convolution. The inter-layer mapping correlation matrix containing the fault risk transmission factor is loaded into the video memory. For each business layer node in the graph structure, all underlying physical nodes supporting business operation are indexed by the matrix. The current node feature vectors of these physical nodes are extracted, and the feature vector of each physical node is multiplied by its corresponding fault risk transmission factor. The weighted feature vectors of all physical nodes are accumulated by vector addition, and the accumulation result is directly injected into the corresponding business unit node to generate vertical risk components, thus completing the bottom-up risk projection of the physical layer hardware status to the business layer logical channel. The horizontal diffusion process involves determining the convolutional receptive field using graph convolution operators based on the topological connectivity defined by the service layer matrix. This process involves locking the service layer network and using the service layer matrix as the adjacency operator. For each service layer node, the set of its first-order neighbor nodes that are directly connected in the logical topology is identified, and this set is the topological receptive field. Mean aggregation is performed on the feature vectors of all neighbor nodes within the receptive field to generate horizontal propagation components, capturing the congestion level of the current service flow on adjacent logical channels and simulating the lateral spread of fault risks within the same network layer. The vertical risk component and the horizontal propagation component are added element-wise, that is, the values ​​of the two vectors at corresponding positions are summed to complete the linear superposition of cross-layer information and same-layer information; the superposition result is input into a nonlinear activation function (ReLU is used in this embodiment) to set all values ​​less than zero in the vector to zero in order to remove noise interference and generate node feature vectors. The multi-channel graph neural network layer adopts a multi-layer stacked deep network structure (in this embodiment, a stacked structure of 2 to 3 layers is preferred to avoid oversmoothing while capturing deep topological information). In this embodiment, the multi-channel graph neural network layer is specifically composed of 3 graph neural network sub-layers with identical structures and independent parameters, sequentially cascaded; each layer is designed with a parallel processing mechanism, including a vertical aggregation channel specifically for handling cross-layer influences and a horizontal diffusion channel specifically for handling same-layer propagation. To ensure that the vertical risk component output by the vertical aggregation channel and the horizontal propagation component output by the horizontal diffusion channel can be fused, a fully connected linear transformation matrix is ​​set to uniformly map the input physical node feature vector and the business node feature vector to a hidden layer feature space of the same dimension; Constructing the dataset required for training the graph neural network model. Collecting operational status data and network topology snapshots of physical layer devices and OSU service layers over historical periods. Specifically, extracting performance metrics (including but not limited to optical power, bit error rate, and CPU utilization) of physical layer nodes in each sampling period as initial features of physical nodes, and extracting operational metrics (including but not limited to transmission latency and packet loss rate) of OSU service layer nodes in the corresponding period as initial features of service nodes. Simultaneously, collecting fault alarm logs recorded in the entire network management system, marking nodes with fault risks as positive samples (risk nodes), and marking normally operating nodes as negative samples (safe nodes), thereby constructing a graph structure sample set containing feature matrices, topology graph structure, and node risk labels. Dividing the graph structure sample set into a training set, validation set, and test set according to a preset ratio of 7:2:1; After generating the node feature vector, the computational model uses the output layer (e.g., a Softmax classifier or a fully connected regression layer) to map the node feature vector to a fault risk prediction result. During the model training phase, an end-to-end supervised learning approach is adopted: the fault risk prediction result output by the model is compared with the true risk labels in the graph structure sample set, and the loss function value is calculated (for risk classification tasks, this embodiment uses the cross-entropy loss function). The gradient of the loss function with respect to each learnable parameter in the network (including the weight matrices of each layer) is calculated using the backpropagation algorithm, and an optimizer (preferably the Adam optimizer in this embodiment) is used to update the network parameters according to the set learning rate. By iterating the above forward feature calculation and backpropagation parameter update process multiple times on the training set and monitoring the model performance on the validation set to prevent overfitting, until the loss function converges, a trained multi-channel graph neural network model is obtained.

[0030] By constructing a dual-channel mechanism that combines vertical aggregation and horizontal diffusion, the physical and logical dimensions of fault evolution are effectively decoupled. Vertical aggregation precisely quantifies the differentiated penetration of underlying hardware into upper-layer business processes using risk transmission factors, while horizontal diffusion keenly captures the lateral congestion characteristics at the business logic level. Specifically, the time slot analysis network associates the extracted spatiotemporal state sequence with the transmission time slots of fgOTN, injects a time sequence identifier corresponding to the transmission period of the OSU service frame, and characterizes the positional relationship of different transmission time slots in the multiframe sequence. Based on the traffic fluctuation correlation analysis mechanism, the performance parameter change pattern of different time slots in the historical transmission cycle is detected, the time sequence anomaly pattern is identified, and the time slot significance weights that reflect the coupling relationship of traffic characteristics between time slots are generated. The time slot saliency weight is used to perform weighted processing on the spatiotemporal state sequence for transmission time slots, and a time slot-aware context vector reflecting the evolution trend of network channel state is extracted. The fault state of the current fgOTN communication link is determined based on the context vector. The process of collecting node feature vectors generated from multiple consecutive time slices based on the time sliding window mechanism and constructing a spatiotemporal state sequence involves creating a first-in-first-out fixed-length queue as a sliding window container; setting the time length parameter of the sliding window (60 time slices in this embodiment); whenever a new node feature vector is generated, it is pushed to the tail of the queue, while the oldest vector at the head of the queue is removed; arranging all the currently lingering node feature vectors in the queue in chronological order and combining them into a matrix structure, which is the spatiotemporal state sequence, recording the state evolution trajectory of network nodes in the most recent continuous time period; The process of associating the extracted spatiotemporal state sequence with the transmission time slots of fgOTN and injecting the timing identifier corresponding to the transmission period of the service frame of the optical service unit is called establishing a time slot indexing mechanism. Each time slot within the sliding window is assigned an integer index between 1 and 20 to mark which specific time slot position in the multiframe period the data corresponding to that time slot was collected from. Then, the integer index is converted into a vector form (one-hot encoded vector in this embodiment) and concatenated into the corresponding node feature vector, so that the model can distinguish the phase relationship of different time slots in the physical transmission period. The process of detecting the performance parameter variation patterns of different time slots within historical transmission cycles and generating time slot significance weights based on the traffic fluctuation correlation analysis mechanism specifically involves: establishing a benchmark model using statistical methods; calculating the mean and variance of features at the same time slot position for several historical multiframe cycles to construct a normal fluctuation range; comparing the feature data of each time slot in the current input sequence with the historical benchmark to calculate the deviation; calculating the absolute value of the difference between the current time slot feature data and the historical feature mean, and dividing the absolute value of the difference by the historical standard deviation to obtain the standardized deviation index; processing the standardized deviation index using the Softmax normalization function, calculating the percentage of the natural index value of the standardized deviation index of the current time slot relative to the sum of the natural index values ​​of all time slots within the sliding window, and determining the percentage value as the significance weight value of the time slot; and generating a weight vector with a length consistent with the sliding window size. The process of extracting the time slot-aware context vector by weighting the spatiotemporal state sequence with time slot saliency weights is to perform element-wise weighting operations. The weight vector generated above is applied to the spatiotemporal state sequence, and the feature vector of each time step in the sequence is multiplied and weighted, thereby amplifying the feature signal of abnormal time slots and suppressing the background noise of stable time slots. The weighted sequence is input into a recurrent neural network layer (in this embodiment, a long short-term memory network is used), which maps the entire variable-length sequence information into a fixed-length hidden state vector, which is the time slot-aware context vector, highly condensing the network operation trend and potential risk characteristics at the current moment. The Long Short-Term Memory (LSTM) network is a bidirectional LSM network with two stacked layers, and the hidden layer dimension is set to 128 to simultaneously capture the contextual dependencies of the multiframe sequence. The fully connected classification layer uses the Softmax function to calculate the probability distribution, and the fault determination threshold is set to 0.8. Furthermore, the baseline model for generating slot saliency weights is calculated based on the statistical deviation of the past 100 historical multiframe cycles to ensure the accuracy of identifying abnormal slots. The process of determining the fault status of the current fine-grained optical transport network communication link based on context vectors specifically involves constructing a fully connected classification layer; inputting the extracted time slot-aware context vector into the fully connected layer containing the Softmax activation function to calculate the probability distribution values ​​of fgOTN in normal and fault states; setting a judgment threshold (0.8 in this embodiment); when the output fault probability value exceeds the threshold, it is determined that the current communication link has an impending fault risk, and the specific prediction result is output. The prediction result is a structured dataset containing fault confidence and fault source location information; wherein, the fault confidence is the probability value exceeding a preset threshold calculated in the previous steps, which quantifies the possibility of fault occurrence; the fault source location information clearly indicates the specific location of the anomaly, including the globally unique identifier of the physical node, the physical port number where the anomaly occurred, and the logical channel identifier of the business unit directly affected by the anomaly. By introducing a correlation analysis mechanism between multiframe timing identifiers and traffic fluctuations, the granularity of fault prediction is refined from macroscopic time slices to microscopic transmission time slots. The characteristic signals of abnormal time slots are effectively amplified by using saliency weights, and background noise of stable periods is suppressed. Combined with LSTM networks, the nonlinear coupling characteristics between time slots are accurately captured. The final output is a structured result containing both physical and logical positioning.

[0031] This invention addresses the strong coupling between physical bearers and logical services in fgOTN networks. By calculating dynamic time slot occupancy rates and converting them into fault risk propagation factors, it establishes a precise cross-layer mapping relationship, effectively quantifying the nonlinear risk penetration of lower-layer hardware congestion into upper-layer OSU services. Furthermore, by utilizing the Laplace decomposition of the joint topology matrix and the vertical / horizontal dual-dimensional feature calculation of graph convolution operators, it overcomes the shortcomings of traditional methods in capturing cross-layer fault propagation and same-layer risk diffusion, achieving a deep fusion analysis of the network's global topology and local spatiotemporal state.

[0032] Example 2: This embodiment applies a topology-based fault prediction method for fgOTN communication networks to an fgOTN communication network scenario that includes 5 physical nodes (labeled P1 to P5, representing the core switch of the control center and the aggregation equipment of 4 stations along the line, respectively) and 3 key OSU services (labeled S1 to S3, representing train operation control signals, platform high-definition monitoring video streams, and internal office dedicated lines, respectively).

[0033] A full scan of the network is performed. Based on the physical optical cable connection, when there is a direct optical path connection between P1 and P2, or between P2 and P3, the corresponding coordinate element in the 5×5 physical layer matrix is ​​set to 1. At the same time, the OSU configuration is parsed. If the S1 service logic needs to be transmitted through the P1, P2, and P3 nodes, the adjacency relationship between S1 and other services if there is a logical channel association is defined in the 3×3 service layer matrix. For the physical port (designated as port P2) carrying the video stream service S1, the system performs in-depth analysis of the fgOTN frame structure and identifies that S1 actually occupies 12 tributary time slots in the payload area during this multiplexing cycle. Given that the total time slot capacity of this port is 20 units, the dynamic time slot occupancy rate is calculated. This value is then mapped using a Sigmoid nonlinear weighting function to generate a fault risk transmission factor, quantifying the sensitivity of the S1 service to the performance fluctuations of the underlying physical port due to its high bandwidth resource consumption. Construct an 8×8 joint topology matrix. Place the physical layer matrix in the upper left region and the service layer matrix in the lower right region. In the upper right off-diagonal block, fill the element values ​​at the intersection of node P2 and service S1 with the calculated fault risk transmission factor, and fill the remaining positions without bearing relationship with 0. Fill the lower left region with its transpose matrix. Calculate the degree matrix of this joint matrix and generate a normalized Laplace matrix. Perform eigenvalue decomposition on this normalized Laplace matrix. Since the total number of nodes in this small-scale application scenario is 8, which is less than the preset spatial embedding dimension of 16 under the large network architecture, in order to fully preserve the global structural features of this local area network, this embodiment actually extracts the eigenvectors corresponding to all 7 smallest non-zero eigenvalues ​​of the matrix, and combines them column by column to form a feature matrix. Each row of data is the topology position identifier of the corresponding physical nodes P1-P5 and service nodes S1-S3 in the entire network architecture with a dimension of 7. A graph structure was established using the joint topology matrix as the framework. Real-time data was collected for the input optical power of node P2 (-12dBm, signal-to-noise ratio 25dB), and the channel latency and jitter of service S1 (4ms, 0.2ms). After normalizing these real-time performance data, they were concatenated with the previously generated topology location identifiers using feature dimensions to form a high-dimensional node attribute vector, which was then assigned to the corresponding node in the graph. At this point, the spatiotemporal state graph not only contained the structural information of P2 supporting S1 but also incorporated the state information that "P2's current optical power is low." Two-dimensional processing is performed using graph convolution operators. In the vertical dimension, the optical power degradation features of node P2 are projected onto service node S1 using the fault risk transmission factor as weight, simulating the penetration of physical faults into the service layer. In the horizontal dimension, service node S1 aggregates the features of its logically adjacent service S2 to determine whether there is congestion in the same layer. The vertically aggregated risk component and the horizontally diffused propagation component are superimposed and activated to generate a node feature vector that integrates cross-layer risk and same-layer impact. The node feature vector sequence was collected over 60 consecutive time slices using a sliding time window, and a time sequence identifier was injected to distinguish whether the data belonged to the 5th or 10th time slot in the multiframe. Based on traffic fluctuation analysis, it was found that the historical variance of the 5th time slot was relatively large, so a high significance weight was automatically generated to amplify the features of this time slot. The weighted sequence was then input into an LSTM network to extract the time slot-aware context vector. Finally, the classifier output result predicted that the probability of port failure of node P2 within the next 10 minutes was 88%.

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting faults in a topology-computing-based fgOTN communication network, characterized in that, include: Collect real-time performance data, physical connection relationships, and OSU configuration of the fgOTN communication network to construct the physical layer matrix and service layer matrix; The dynamic timeslot occupancy rate of OSU services on physical ports is calculated based on the time-division multiplexing frame structure of fgOTN and converted into a fault risk transmission factor. The fault risk transmission factor is used as an element value to construct an inter-layer mapping correlation matrix. A joint topology matrix is ​​established by combining the physical layer matrix and the service layer matrix. Laplace decomposition is performed on the joint topology matrix to generate topology location identifiers. The specific process of calculating the dynamic timeslot occupancy rate of OSU services on physical ports and converting it into a fault risk transmission factor based on the time-division multiplexing frame structure of fgOTN is to construct an index mapping table between physical ports and OSU services based on the OSU configuration. Based on the time-division multiplexing frame structure of fgOTN, the time slot allocation bitmap of the physical port in a single multiplexing cycle is parsed. Based on the time slot allocation bitmap, the actual location and number of tributary time slots occupied by each OSU service carried on the physical port are identified. The ratio of the number of tributary time slots to the total time slot capacity of the physical port frame structure payload area is calculated to generate a dynamic time slot occupancy rate. The dynamic time slot occupancy rate is mapped to a fault risk transmission factor through a nonlinear weighting function. A spatiotemporal state graph is constructed based on topological location identifiers; based on the spatiotemporal state graph, graph convolution operators are used to perform topological feature calculations to generate node feature vectors; the topological feature calculations include vertical and horizontal dimensions, the vertical dimension projects the inter-layer mapping correlation matrix to the OSU service layer by weighting the fault risk transmission factor, and the horizontal dimension defines the topological receptive field based on the service layer matrix and aggregates the neighborhood structure distribution features. The node feature vectors generated by collecting multiple consecutive time slices based on the time sliding window mechanism are used to construct a spatiotemporal state sequence and input into the time slot analysis network to generate the fgOTN communication network fault prediction result.

2. The method for fault prediction in fgOTN communication networks based on topology calculation according to claim 1, characterized in that, The physical layer matrix is a physical adjacency matrix based on the physical connection relationship of the fgOTN communication network, and has dimensions of wherein is the number of physical nodes, and the matrix elements are determined based on the physical connection relationship, with an element value of 1 when there is a physical link between two physical nodes, and 0 otherwise. The service layer matrix is a logical adjacency matrix representing the logical connection relationship of the OSU service logic, and the dimension is wherein is the number of service nodes, and the matrix element is determined based on the OSU configuration. The element value is 1 when there is an OSU logical channel between two service nodes, otherwise 0.

3. The method for fault prediction in fgOTN communication networks based on topology calculation according to claim 1, characterized in that, The interlayer mapping correlation matrix is a heterogeneous correlation matrix with dimensions wherein is the number of physical nodes, is the number of service nodes; The element values ​​in the inter-layer mapping association matrix are the fault risk transmission factors at the corresponding intersection positions of physical nodes and service nodes that have a bearing relationship in the inter-layer mapping association matrix. For physical nodes and service nodes that do not have a bearing relationship, the element value at the corresponding intersection position in the inter-layer mapping association matrix is ​​0; A joint topology matrix is ​​constructed using a block matrix structure. The physical layer matrix is ​​located in the upper left diagonal block of the joint topology matrix, and the service layer matrix is ​​located in the lower right diagonal block of the joint topology matrix. The inter-layer mapping association matrix is ​​located in the upper right off-diagonal block of the joint topology matrix, and the transpose of the inter-layer mapping association matrix is ​​located in the lower left off-diagonal block of the joint topology matrix.

4. The method for fault prediction in fgOTN communication networks based on topology calculation according to claim 1, characterized in that, The Laplace decomposition is performed to determine the connection density of network nodes based on the joint topology matrix, generate connection metric parameters that reflect the hub status of nodes in the network, and use the connection metric parameters to standardize the network connection relationships to establish an association model that reflects the global connection architecture of the fgOTN network. Structural features are extracted from the association model, the connection patterns and link coupling relationships of network nodes are analyzed, and topological location identifiers that identify the distribution characteristics of each physical node and service node in the network topology are generated according to the order of their impact on network connectivity.

5. The method for fault prediction in fgOTN communication networks based on topology calculation according to claim 1, characterized in that, The specific process of constructing a spatiotemporal state graph based on topological location identifiers involves establishing a topological skeleton of the graph by using a joint topological matrix as the edge set and physical nodes and service nodes as the node set; concatenating the topological location identifiers with the real-time performance data by feature dimensions to generate node attribute vectors; the real-time performance data includes input optical power, output optical power, optical signal-to-noise ratio, and bit error rate for physical nodes, and OSU channel latency, latency jitter, and OSU frame loss rate for service nodes. The node attribute vectors are assigned to the corresponding nodes in the node set of the graph to construct a spatiotemporal state graph.

6. The method for fault prediction in fgOTN communication networks based on topology calculation according to claim 1, characterized in that, The use of graph convolution operators to perform topology-based feature computation specifically includes vertical aggregation and horizontal diffusion. The vertical aggregation is achieved by using the graph convolution operator with the inter-layer mapping correlation matrix as the adjacency weight matrix for cross-layer convolution, and projecting the state features of the physical layer nodes to the corresponding OSU service layer nodes according to the fault risk transmission factor; the horizontal diffusion is achieved by using the graph convolution operator to determine the convolution receptive field based on the topological connection relationship defined by the service layer matrix, and performing neighborhood feature aggregation on the OSU service layer nodes.

7. The method for fault prediction in fgOTN communication networks based on topology calculation according to claim 1, characterized in that, Specifically, the time slot analysis network associates the extracted spatiotemporal state sequence with the transmission time slots of fgOTN, injects a time sequence identifier corresponding to the transmission period of the OSU service frame, and characterizes the positional relationship of different transmission time slots in the multiframe sequence. Based on the traffic fluctuation correlation analysis mechanism, the performance parameter change pattern of different time slots in the historical transmission cycle is detected, the time sequence anomaly pattern is identified, and the time slot significance weights that reflect the coupling relationship of traffic characteristics between time slots are generated. The spatiotemporal state sequence is weighted according to the transmission time slot using the time slot significance weight, and a time slot-aware context vector reflecting the evolution trend of network channel state is extracted. The fault state of the current fgOTN communication link is determined based on the time slot-aware context vector.

Images