Fault positioning method and system for telecommunication network

Abstract

The invention discloses a fault positioning method and system for a telecommunication network. The method comprises the following steps: initializing a topology time causal graph; the convolution operation of the topological domain and the convolution operation of the time domain are combined into combined convolution operation of the topological domain and the time domain; self-adaptive embedding is defined based on GCN, and event types and initial causal structure information of physical node topology are injected into event sequence embedding; constructing a topological time causal transformer model; optimizing model parameters of the topological time causal transformer through a maximum log-likelihood function; and iteratively updating the causal weight, and finally storing the causal weight as a causal graph between event types. The method has the beneficial effects that the optimal causal relationship structure can be obtained by abstracting the network alarm time into the alarm event type, the physical node to which the event belongs and the event sequence represented by the event occurrence time, and meanwhile, the corresponding topological time causal transformer model is constructed, so that the global dependency can be captured; and finally, storing a causal graph between event types, and further realizing root cause positioning.

Description

Technical Field

[0001] The present application relates to the field of network operation and maintenance, and in particular to a method and system for locating faults in a telecommunication network. Background Art

[0002] At present, telecommunication network refers to a system that realizes data and voice transmission between mobile devices or between mobile devices and fixed networks through wireless communication technology. It includes various network devices such as base stations, mobile switching centers, routers, etc., which are interconnected through physical links and radio waves to provide users with a wide range of communication services. In a telecommunication network, network devices are interconnected through physical links. When a device fails, it may affect other devices connected to it, trigger alarms, and even cause service interruptions in severe cases. With the expansion of network scale and exponential growth in complexity, the frequency of network failures has also increased dramatically, seriously affecting network quality and user experience. Therefore, how to effectively detect, diagnose and repair network failures has become the key to ensuring the stable operation of telecommunication networks.

[0003] Fault location in telecommunication networks is usually identified through alarms. A fault event sequence in a network consists of discrete events in continuous time, which often appear in the form of abnormal alarms in communication network scenarios. Based on the historical fault event sequence, we can find the causal relationship between network alarm types, which helps to locate the root cause of the alarm efficiently and accurately, thereby improving operation and maintenance efficiency and reducing operation and maintenance costs. However, there are a large number of redundant alarms in the wireless network alarm information, which poses a challenge to the accurate location of the root cause. Since the devices in the mobile communication base station are physically or logically connected, when a device fails, a large number of alarm information directly related to the fault will be generated. At the same time, other devices associated with the device may also be affected, generating a large amount of redundant alarm information that is not directly related to the fault. In addition, the characteristics of the alarm event sequence, such as sparse data, small amount of information, and low time accuracy, pose many challenges to the existing root cause location methods. Summary of the invention

[0004] The main purpose of the present application is to provide a method and system for locating faults in a telecommunications network, so as to display alarm information through a cause-and-effect diagram, thereby quickly finding the root cause of the fault.

[0005] In order to achieve the above object, the present application provides: a fault location method for a telecommunications network, the method comprising:

[0006] Initialize the topological temporal causal graph based on the topological structure and time steps of the event sequence;

[0007] The convolution operations in the topological domain and the time domain are combined into a joint convolution operation in the topological time domain, and a topological time causal graph is generated;

[0008] Define adaptive embedding based on GCN (graph convolutional network) to inject the initial causal structure information of event type and physical node topology into the event sequence embedding;

[0009] Build a topological temporal causal transformer model and capture global dependencies;

[0010] Optimizing the topological temporal causal transformer model parameters by maximizing a log-likelihood function;

[0011] The causal weights are updated iteratively and finally saved as a causal graph between event types to locate the fault.

[0012] Furthermore, the method of combining the convolution operations in the topological domain and the time domain into a joint convolution operation in the topological time domain and generating a topological time causal graph includes:

[0013] Let B∈R N×N is a weighted causal matrix. Given an event type n,n′∈N, B n,n′ Characterize the causal strength from type n to type n′, let (n,v,t) and (n′,v′,t′) be the events observed in the event sequence E, where Calculate B n,n′ The topological time causal graph is convoluted and defined as follows:

[0014]

[0015] Among them, φ (v,t),(v′,t′) is the convolution kernel of the topological domain, is the convolution kernel of the time domain, n, n′ is the given event type, G V represents the topological structure in the network, and T represents the dimension of time.

[0016] Furthermore, the method of defining adaptive embedding based on GCN (graph convolutional network) and injecting initial causal structure information of event type and physical node topology into event sequence embedding includes:

[0017] Build an embedding layer based on GCN and inject the topological, causal, and temporal information of event observations;

[0018] Adaptive encoding based on GCN, specifically:

[0019] Physical node encoding, let Represents the vertex subsequence of the alarm event, initializes the physical node embedding matrix D V ∈R M×|V| , where D V The jth column represents the vertex v j M-dimensional embedding, assuming is the input of the kth layer of GCN, where At this time, the topological graph convolution is defined as inputting the topological relationship of the physical nodes into the node embedding. At this time, the topological graph convolution is:

[0020]

[0021] Where W V is the parameter matrix, matrix A is the fixed topology structure, is the graph structure learned dynamically by parameterizing A;

[0022] Modify the embedding of the event vertex subsequence to:

[0023]

[0024] in is the one-hot encoding matrix of the event vertex subsequence, is the vertex v j A one-hot encoding of , j∈V;

[0025] Event type code, let Represents the type subsequence of the alarm event. The event types in N are mapped to high-dimensional features using a trainable linear projection. The embedding matrix of the event type is represented as D N ∈R M×|N| , where M is the embedding dimension and the matrix D N The jth column corresponds to event type n j The M-dimensional embedding matrix D N Convert discrete event types into continuous real-valued vectors;

[0026] Construct a causal graph convolution to fuse the causal semantic information into the event type embedding matrix to obtain the feature representation of event types at the causal level. represents the input of the kth layer of GCN, where Define semantic graph convolution as capturing patterns in causal graphs:

[0027]

[0028] in is the parameterized representation of the trainable matrix B, σ is the activation function, W N The parameter matrix semantic graph convolution is defined to capture the intrinsic pattern of the causal graph; after the training is stable, Infer the most appropriate causal graph structure for the event type and reshape the embedding of the event type subsequence into:

[0029]

[0030] in is the one-hot encoding matrix of the event type subsequence, the vector Is the event type n j A one-hot encoding of , j∈N;

[0031] Time code, order Initialize the learnable time embedding matrix D for the time subsequence of the alarm event T ∈R M×|T| , where D T The jth column of represents the time step t j The M-dimensional embedding of the alarm event time subsequence E T is encoded into an M×|T|-dimensional matrix X t′ .

[0032] Furthermore, the method of constructing a topological temporal causal transformer model and capturing global dependencies includes:

[0033] A topological transformer model is constructed, including a multi-head self-attention layer (MHSA) and a fully connected feedforward layer (FFN). The node embedding features are input into the topological transformer, and the multi-head self-attention layer (MHSA) and the fully connected feedforward layer (FFN) are used to sequentially integrate the information of all alarm times.

[0034] Constructing a causal transformer model, fusing the output of the topological transformer model and event type embedding to encode high-dimensional features for each event;

[0035] A temporal transformer model is constructed to encode the temporal information of alarms and capture the global contextual relations.

[0036] Further, after constructing the topological temporal causal transformer model and capturing the global dependency, the method further includes:

[0037] constructing a continuous-time conditional intensity function based on the output of the topological-temporal causal transformer model;

[0038] The log-likelihood maximization function is constructed through the continuous-time conditional intensity function.

[0039] The present invention further discloses a telecommunication network fault location system, comprising:

[0040] An initialization module, used to initialize the topological temporal causal graph based on the topological structure and time steps of the event sequence;

[0041] A joint operation module, used for combining the convolution operations in the topological domain and the time domain into a joint convolution operation in the topological time domain, and generating a topological time causal graph;

[0042] The information injection module is used to define adaptive embedding based on GCN (graph convolutional network) to inject the initial causal structure information of event type and physical node topology into the event sequence embedding;

[0043] Model building module, used to build topological temporal causal transformer models and capture global dependencies; also used to,

[0044] The topological temporal causal transformer model parameters are optimized by maximizing the log-likelihood function; and further used for,

[0045] The causal weights are updated iteratively and finally saved as a causal graph between event types to locate the fault.

[0046] Furthermore, the model building module is also used to:

[0047] constructing a continuous-time conditional intensity function based on the output of the topological-temporal causal transformer model;

[0048] The log-likelihood maximization function is constructed through the continuous-time conditional intensity function.

[0049] The present invention further discloses a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned telecommunication network fault location method is implemented.

[0050] The present invention further discloses an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned telecommunication network fault location method when executing the program.

[0051] The present invention further discloses a computer program product, comprising a computer program, wherein the computer program implements the above-mentioned telecommunication network fault location method when executed by a processor.

[0052] Compared with the prior art, the present invention provides a telecommunication network fault location method and system, which have the following beneficial effects: by abstracting the network alarm time into an event sequence represented by the alarm event type, the physical node to which the event belongs, and the event occurrence time, the optimal causal relationship structure between multiple alarm types can be discovered, and a corresponding topological time causal transformer model is constructed at the same time, which can capture global dependencies and finally save them as a causal graph between event types, thereby realizing root cause location. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The drawings constituting a part of this application are used to provide a further understanding of this application, so that other features, purposes and advantages of this application become more obvious. The schematic implementation drawings of this application and their descriptions are used to explain this application and do not constitute an improper limitation on this application. In the drawings:

[0054] Figure 1 This is a schematic diagram of the process of the fault location method of the telecommunication network according to the present embodiment;

[0055] Figure 2 Schematic diagram of constructing a topological temporal causal transformer model in this implementation mode;

[0056] Figure 3 This is a schematic diagram of the topological transformer model architecture of this implementation method;

[0057] Figure 4 This is a schematic diagram of the causal transformer model architecture of this embodiment;

[0058] Figure 5 This is a schematic diagram of the architecture of the time transformer model in this implementation mode;

[0059] Figure 6 This is a schematic diagram of the structure of a fault location system for a telecommunication network according to this embodiment;

[0060] Figure 7 Schematic diagram of comparison results of accuracy, recall and F1 curve of the training performance of TTCTH in this embodiment on a real data set;

[0061] Figure 8 Schematic diagram of the comparison of the accuracy results of TTCTH in this embodiment and the comparison method on two real-world data sets;

[0062] Fig. 9 Schematic diagram of comparison of recall results of TTCTH in this embodiment and comparison method on two real-world datasets;

[0063] Fig.10 Schematic diagram of the comparison of F1 results of TTCTH in this embodiment and the comparison method on two real-world datasets;

[0064] Fig.11 A schematic diagram showing the comparison of the accuracy results of the positioning method disclosed in this embodiment on two real data sets;

[0065] Fig.12 A schematic diagram showing comparison of recall rate results of the positioning method disclosed in this embodiment on two real data sets;

[0066] Fig.13 It is a schematic diagram of the comparison of the F1 results of the positioning method disclosed in this embodiment on two real data sets. DETAILED DESCRIPTION

[0067] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the implementation mode of the present application will be clearly and completely described below in conjunction with the drawings in the implementation mode of the present application. Obviously, the described implementation mode is only a part of the implementation mode of the present application, not all the implementation modes. Based on the implementation mode in the present application, all other implementation modes obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present application.

[0068] It should be noted that the terms "first", "second", etc. in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0069] In the present application, the terms "upper", "lower", "left", "right", "front", "back", "top", "bottom", "inner", "outer", "middle", "vertical", "horizontal", "lateral", "longitudinal" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings. These terms are mainly used to better describe the present application and its implementation methods, and are not used to limit the indicated devices, elements or components to have a specific orientation, or to be constructed and operated in a specific orientation.

[0070] In addition, some of the above terms may be used to express other meanings in addition to indicating orientation or positional relationship. For example, the term "on" may also be used to express a certain dependency or connection relationship in some cases. For those of ordinary skill in the art, the specific meanings of these terms in this application can be understood according to specific circumstances.

[0071] In addition, the term "plurality" shall mean two or more.

[0072] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0073] In communication networks, abnormal phenomena are usually identified through alarms. Due to the scale of the network and the interconnected structure within the system, administrators may face millions of alarms every day. A single fault in the network will trigger a large number of alarms of various types on multiple connected devices. It is very difficult for administrators to quickly locate the root cause of the fault from the alarm storm. In this case, using a cause-and-effect diagram to display the alarm information can help administrators quickly find the root cause of the fault. Communication network alarm events can be abstracted into event sequences represented by the alarm event type, the physical node to which the event belongs, and the time when the event occurred. Therefore, the fault alarm records collected over a period of time can be represented by event sequences. Formal representation. Where: N represents the set of event types; V represents the set of vertices in the topological structure; T represents the recording time of the event sequence; (n i ,v i ,t i ) represents the i-th event type n i , at t i The moment v occurs in the topological structure i Vertex. Let the undirected graph G v =(N,E v ) is used to represent the topological structure diagram between physical nodes V, E v is the edge set in the graph. Since the alarm event sequence occurs on the physical nodes in the topological network, an alarm event is not only stimulated or inhibited by the cause event occurring in the sequence, but also stimulated or inhibited by the cause event in the topological neighbor. Therefore, given a set of observed alarm event sequences n i ∈N,ν i ∈V,t i ∈T, the vertex topology graph G corresponding to the event V and the initial causal structure G of the event type N , thereby finding the optimal causal relationship structure G between N alarm types N , and assist in locating the root cause. N =(N,E N ) is defined as a directed acyclic graph, which is used to characterize the causal relationship between event types N, where N represents the number of alarm types, E N is the set of edges in the causal graph.

[0074] refer to Figure 1 , is a flow chart of a method for locating a fault in a telecommunication network in this implementation method, that is, a method for locating a fault in a telecommunication network in this implementation method, the method comprising:

[0075] S100: Initialize the topological temporal causal graph based on the topological structure and time steps of the event sequence;

[0076] Alarm events are triggered by their topological neighbors through different paths. Therefore, the GCN method with topological time is used to learn the initial causal graph according to the input event sequence and its topological structure. The short-term causal relationship between event types can be captured by exploring the spatial dependency of neighbors and the temporal dependency of the nearest time. Specifically:

[0077] Let B∈R N×N is a weighted causal matrix. Given an event type n,n′∈N, B n,n′ Characterize the causal strength from type n to type n′, let (n,v,t) and (n′,v′,t′) be the events observed in the event sequence E, where Calculate B n,n′ The topological time causal graph is convoluted and defined as follows:

[0078]

[0079] Among them, φ (v,t),(v′,v′) is the convolution kernel of the topological domain, is the convolution kernel of the time domain, n, n′ is the given event type, G V represents the topological structure in the network, and T represents the dimension of time.

[0080] Furthermore, the function φ (v,t),(v′,t′) Summarizes the accidental effects of its topological neighbors. Accumulating the accidental effects of historical events, it can be seen that the topological structure remains unchanged throughout time. Therefore, the convolution definition formula of the topological time causal graph can be rewritten as: When the widely used index form is chosen When, the hyperparameter characterizes the time decay random effect ∈ n,n′ and At this time, the convolution kernel in the time domain can be expressed as: Where A is G V The adjacency matrix, A v,v′ =1, is the k-hop adjacency matrix, θ v,v′,k represents the accidental influence transmitted from node V to node v′ through the k-hop path.

[0081] S200: combining the convolution operations in the topological domain and the time domain into a joint convolution operation in the topological time domain, and generating a topological time causal graph.

[0082] S300: Define adaptive embedding based on GCN (graph convolutional network) to inject the initial causal structure information of event type and physical node topology into event sequence embedding; specifically:

[0083] Build an embedding layer based on GCN and inject the topological, causal, and temporal information of event observations;

[0084] Adaptive encoding based on GCN, specifically:

[0085] Physical node encoding, let Represents the vertex subsequence of the alarm event, initializes the physical node embedding matrix D V ∈R M×|V| , where D V The jth column represents the vertex v j M-dimensional embedding, assuming is the input of the kth layer of GCN, where At this time, the topological graph convolution is defined as inputting the topological relationship of the physical nodes into the node embedding. At this time, the topological graph convolution is:

[0086]

[0087] Where W V is the parameter matrix, matrix A is the fixed topology structure, is the graph structure learned dynamically by parameterizing A;

[0088] Modify the embedding of the event vertex subsequence to:

[0089]

[0090] in is the one-hot encoding matrix of the event vertex subsequence, is the vertex v j A one-hot encoding of , j∈V;

[0091] Event type code, let Represents the type subsequence of the alarm event. The event types in N are mapped to high-dimensional features using a trainable linear projection. The embedding matrix of the event type is represented as D N ∈R M×|N| , where M is the embedding dimension and the matrix D N The jth column corresponds to event type n j The M-dimensional embedding matrix D N Convert discrete event types into continuous real-valued vectors;

[0092] Construct a causal graph convolution to fuse the causal semantic information into the event type embedding matrix to obtain the feature representation of event types at the causal level. represents the input of the kth layer of GCN, where Define semantic graph convolution as capturing patterns in causal graphs:

[0093]

[0094] in is the parameterized representation of the trainable matrix B, σ is the activation function, W N The parameter matrix semantic graph convolution is defined to capture the intrinsic pattern of the causal graph; after the training is stable, Infer the most appropriate causal graph structure for the event type and reshape the embedding of the event type subsequence into:

[0095]

[0096] in is the one-hot encoding matrix of the event type subsequence, the vector Is the event type n j A one-hot encoding of , j∈N;

[0097] Time code, order Initialize the learnable time embedding matrix D for the time subsequence of the alarm event T ∈R M×|T| , where D T The jth column of represents the time step t j The M-dimensional embedding of the alarm event time subsequence E T is encoded into an M×|T|-dimensional matrix X t′ .

[0098] S400: Build a topological temporal causal transformer model and capture global dependencies.

[0099] refer to Figure 2-Figure 5 ,After the embedding layer merges the information of causal and topological graphs into the event sequence, the topological temporal causal transformer is designed to jointly model the global and long-term dependencies between event types inspired by topological neighbors. It consists of three separable transformer blocks: topological transformer model, causal transformer model and temporal transformer model. The construction methods for the topological transformer model, causal transformer model and temporal transformer model are as follows:

[0100] A topological transformer model is constructed to encode high-dimensional features to capture hidden spatial dependencies that evolve over time, including a multi-head self-attention layer (MHSA) and a fully connected feed-forward layer (FFN) to embed nodes into feature X s Input into the topological transformer, the multi-head self-attention layer (MHSA) and the fully connected feed-forward layer (FFN) are used to sequentially integrate the information of all alarm times. Among them, the multi-head self-attention (MHSA) mechanism can learn multiple spatial dependency patterns. Each head uses scaled dot product attention in parallel to process a spatial dependency pattern. The spatial attention S s The calculation method is:

[0101]

[0102] in and They are the query matrix, key matrix, and value matrix respectively. and Q s , K s and V s Dimension. s , K s and V are matrices consisting of node embedding features X s With the weight matrix and The linear transformation of is calculated. Let X s =X s′ .therefore,

[0103] Assume that different attention outputs are regarded as MHSA. The MHSA output is the concatenation of the outputs of h attention heads, as shown below:

[0104]

[0105] in is the aggregation matrix. The attention output MHSA(X s ) is then fed through a two-layer point-wise feed-forward neural network. The transformation here applies residual connections around each module followed by layer normalization. The calculation formula for the entire FFN (feed-forward neural network) process is:

[0106]

[0107] Generate a hidden representation R of the input event sequence s′ , specifically:

[0108] LN(X s )=LayerNorm(MHSA(X s )+X s ),

[0109] R s′ =LayerNorm(LN(X s )+FFN(X s ))+X s .

[0110] in are the parameters of the neural network.

[0111] Construct a causal transformer model, fuse the output of the topological transformer model and the event type embedding, encode high-dimensional features for each time; embed the event type of the event sequence into Xc with the hidden representation R from the topology transformer block c′ Added to the causal transformer model, the multi-head self-attention mechanism is used to model causal dependencies, causal attention S c The calculation formula is:

[0112]

[0113] in and and make are the different attention outputs of h attention heads. The MHSA output is the concatenation of the outputs of h attention heads, as shown below: in is an aggregation matrix. Then, the attention output MHSA(X c ) is fed into a positional feedforward neural network to generate a hidden representation R of the input event sequence c′ , we get the formula:

[0114] R c′ =LayerNorm(LN(X c )+FFN(X c ))+X c ,

[0115] It follows that:

[0116] LN(X c )=LayerNorm(MHSA(X c )+X c ),

[0117]

[0118] in and are the parameters of the neural network. The matrix R c′ ∈R L×M Contains the hidden representation of all event types in the event sequence, where each row corresponds to a specific event type.

[0119] Construct a time transformer model to encode the time information of the alarm and capture the global contextual relationship. Specifically:

[0120] After encoding the temporal embedding layer, the timestamped encoded representation of the event sequence is X t The hidden representation R of the causal transformer layer output c Add them together as input to the time transformer:

[0121] X t =X t′+R c′ ,

[0122] The multi-head self-attention mechanism is used to model temporal dependencies. Calculate the temporal attention S t :

[0123]

[0124] in and and

[0125] make For different attention outputs, the MHSA output is the concatenation of the h attention head outputs:

[0126]

[0127] in is an aggregation matrix.

[0128] Attention Output MHSA(X t′ ) is then passed through a position feed-forward neural network:

[0129] LN(X t )=LayerNorm(MHSA(X t )+X t ),

[0130] Then generate the hidden representation R of the input event sequence t′ :

[0131]

[0132] in and are the parameters of the neural network.

[0133] Matrix R t′ ∈R L×M Contains a hidden representation of all event types, where each row corresponds to a specific event type:

[0134] R t′ =LayerNorm(LN(X t )+FFN(X t ))+X t .

[0135] The output of the Temporal Transformer block is then passed to the recurrent neural network LSTM-CNN:

[0136] R t =conv1.LSTM(R t′ ) / .

[0137] Where R t is the last output of three consecutive transformer model blocks. The addition of LSTM-CNN to the temporal transformer model block is to enhance the model's ability to learn long-term dependencies in sequences.

[0138] After constructing the topological temporal causal transformer model and capturing global dependencies, the fault localization method also includes:

[0139] constructing a continuous-time conditional intensity function based on the output of the topological-temporal causal transformer model;

[0140] The log-likelihood maximization function is constructed through the continuous-time conditional intensity function.

[0141] Specific:

[0142] According to the R obtained in step S400 t , construct the continuous time conditional intensity function λ(t|R t ):

[0143]

[0144] Among them, λ k (t|R t ) is the conditional intensity function of event type k at time t, k∈N, t∈[t j ,t j+1 ].item Represents two consecutive timestamps t j and t j+1 The impact of events in Indicates the impact of historical events on current events, while b k,v represents the probability of event k occurring without considering historical information. To further improve the flexibility of the model, the softplus function is expanded to:

[0145]

[0146] Among them, ξ k,v and π k,v Is a parameter. Based on the sequence The conditional intensity function λ(t|R t ), its log-likelihood function is:

[0147]

[0148] Among them, λ(t|R t ) is defined as follows:

[0149]

[0150] S500: Optimizing the topological temporal causal transformer model parameters by maximizing the log-likelihood function.

[0151] Specifically, in order to ensure that the causal diagram The directed acyclic property of Introduce a smooth constraint:

[0152]

[0153] Where e represents the matrix exponent and ⊙ represents the Hadamard product. Subject to the above non-circularity constraint, the continuous optimization problem is defined as:

[0154]

[0155] The parameters include and Divide For all other parameters except δ, the log-likelihood is maximized to optimize the parameters and δ. The above optimization problem is solved by the augmented Lagrangian method. The Lagrangian function is defined as:

[0156]

[0157] where β>1 and τ<1 are hyperparameters.

[0158] S600: Iteratively update the causal weights, and finally save them as a causal graph between event types, and locate the fault.

[0159] This embodiment further verifies the fault location method of the telecommunication network implemented by the above steps through a real alarm data set, and uses a variety of commonly used causal discovery methods as comparative experiments. The method is to infer the causal graph between the alarm event types based on the real historical alarm data set to compare with the real causal graph. For the comparative experiment, an ablation study is used to verify the performance of each module in the method, and the accuracy, recall rate and F1 score are used as evaluation indicators. The real alarm data set used contains historical alarm data, a topological graph between devices and a real causal graph. The format of the historical alarm data is [warning number, device number, start time, end time], where the warning number represents the alarm type and the device number represents the device. The topological graph describes the network topology between devices and is represented by an undirected graph with devices as nodes. If two devices are connected to each other, there is an edge in the topological graph; otherwise, there is no edge. The real causal graph represents the causal relationship between alarm types. It is a directed graph with alarm types as nodes. Each directed edge connects a pair of ordered alarm types. Exemplary:

[0160] (1) 24V_439N_Microwave dataset: contains 64,598 alarm records of 439 devices and 24 alarm types. The topology graph is represented by a 439*439 symmetric adjacency matrix, and the causal graph is represented by a 24*24 directed matrix.

[0161] (2) 25V_474N_Microwave dataset: contains 48,572 alarm records from 474 devices, with 25 alarm types. The asymmetric adjacency matrix of size 474*474 corresponds to the topological graph, and the directed matrix of size 25*25 corresponds to the causal graph.

[0162] The fault location method for telecommunication networks disclosed in this embodiment can be decomposed into TTGCN and TTCTH. Specifically: TTGCN is a topological time graph convolution block, which is mainly used to capture short-term causal relationships between event types. TTCTH combines the TTGCN block with three stacked multi-head converter blocks to capture long-term causal dependencies hidden in the data set. We compare TTGCN and TTCTH with competitive causal discovery methods, such as NOTEARS, GOLEM, ICALINGAM, DAG-GNN, PC, GES, GAE, and THP.

[0163] NOTEARS: Nonlinear Optimal Transformations for Embedding and Restoring Structure (NOTEARS) is a DAG learning method designed to estimate causal relationships in linear data models. It finds the optimal directed acyclic graph by optimizing a non-convex objective function.

[0164] GOLEM: Causal Discovery Beyond Linear Structural Equation Models (GOLEM) is an enhanced version of the NOTEARS method that introduces some optimization techniques to reduce the number of iterations and find the optimal directed acyclic graph faster.

[0165] ICALINGAM: Independent Component Analysis-based Linear Non-Gaussian Acyclic Model (ICALINGAM) is a causal inference method for discovering causal relationships between variables from data. It infers causal structures by analyzing the non-Gaussianity and linearity of data.

[0166] DAG-GNN: DAG structure learning based on graph neural network (DAG-GNN) is a graph neural network method for processing directed acyclic graph data structures. It uses graph neural networks to predict causal relationships between variables.

[0167] PC: Peter and Clark method (PC) is a classic causal discovery method for constructing a directed acyclic graph between variables. This method infers causal relationships by testing the conditional independence between variables in the data.

[0168] GES: Greedy Equivalence Search (GES) is a classic greedy equivalence search method for learning Bayesian network structures from data. It explores the optimal DAG structure identification through a two-stage greedy search method.

[0169] GAE: Graph Autoencoder (GAE) is used to learn representations of graph-structured data to model nonlinear causal relationships. This method uses an autoencoder to learn a low-dimensional representation of the graph to infer causal relationships.

[0170] THP: Topological Hawkes Process (THP) is a causal structure learning method for modeling event sequences based on Hawkes processes and topological structure information. The method infers the causal graph by maximizing the likelihood of THP based on graph convolution.

[0171] The fault location method disclosed in this embodiment is compared with other methods. Referring to Table 1 and Table 2, the experimental results of the model constructed in this embodiment and other models on the above two exemplary real data sets are shown. It is easy to see that based on the F1 evaluation, the method of this embodiment is significantly better than other methods. In particular, the method of this embodiment is significantly better than all the comparison methods in terms of recall rate and F1 score, while THP is competitive in terms of accuracy. Given that the method of this embodiment and THP take into account topology information, both of them are competitive with other methods.

[0172] The proposed method utilizes stacked transformer model blocks to capture the long-term causal strength hidden in the dataset, although it introduces some redundant causal edges. In this case, THP can easily overlook important true causal information. In contrast, the proposed method is able to find more valuable true causal relationships, based on which operators can effectively explore the root cause localization in network fault management.

[0173] Table 1: Comparison of results of different comparison models on dataset 1

[0174]

[0175]

[0176] Table 2: Comparison of results of different comparison models on dataset 2

[0177]

[0178] refer to Figure 7,For the training performance of TTCTH, the figure shows that the accuracy, recall and F1 evaluation of TTCTH on two real datasets have similar trends. The accuracy score drops relatively little at the beginning, but the recall score rises accordingly. Therefore, the overall F1 score tends to rise. The reason behind this is that the transformer model mechanism captures more long-term causal edges at the cost of introducing small redundant edges.

[0179] Regarding the model stability in this embodiment, in order to avoid the inability to cover all event types due to too small samples, relatively more samples were used in the experiment, and the number of samples was evenly increased on two real data sets, as shown in Table 3.

[0180] Table 3: Introduction of the dataset after increasing the number of samples

[0181]

[0182] As the number of samples increases, Figure 8 , Fig. 9 and Fig.10 We can see that the performance of GOLEM, NOTEARS, GES, DAG-GNN, PC, and GAE on both datasets does not change with the increase in the number of samples. This indicates that these six methods lose the ability to capture long-term causal dependencies. THP shows an upward trend on the 24V_439N_Microwave dataset. However, THP does not have any upward ability on the 25V_474N_Microwave dataset. This indicates that the performance of THP on different real datasets is not stable. Although the precision, recall, and F1 scores of ICALINGAM show an upward trend, these scores are significantly lower than the methods proposed in this embodiment.

[0183] We found that the method proposed in this embodiment performs best compared to the comparison methods. It is worth noting that the increase in samples has a positive impact on the performance of the proposed method on two real-world datasets. These findings show that the method proposed in this embodiment has good robustness in different real-world datasets and outperforms other methods. It is able to capture more long-term causal relationships by increasing continuous samples.

[0184] This embodiment further implements an ablation study to test the sensitivity of the method in this embodiment to sample size and topology, see Figure 11-13, TTCTH(oh) combines the TTGCN block with three stacked single-head transformer blocks. TTCTH(mh) combines the TTGCN block with three stacked multi-head transformer blocks. We see that due to the complexity of real-world datasets, the results of these methods have different trends on the 24V_439N_Microwave and 25V_474N_Microwave datasets.

[0185] from Fig.11 , Fig.12 and Fig.13 It is not difficult to see that the F1 scores of TTGCN, TTCTH(oh), and TTCTH(mh) gradually improve with the increase of samples on both datasets. TTCTH(mh) performs comparable to TTCTH(oh) in accuracy evaluation. However, TTCTH(mh) outperforms TTCTH(oh) in recall and F1 evaluation. Therefore, multi-head TTCTH is able to capture more global causal information than single-head TTCTH because TTCTH(mh) allows the model to jointly obtain causal information from different representation subspaces.

[0186] In summary, this embodiment creatively designs a topological temporal graph convolution to construct an initial causal graph. The combination of topological graph convolution and temporal convolution jointly summarizes the causal influence of its topological neighbors across time through different paths. This graph captures the short-term dependencies between event types. At the same time, a semantic GCN-based embedding is proposed. Semantic causal graph convolution and topological graph convolution aim to extend traditional graph convolution, introduce learnable causal and topological information into event sequence embedding, and accurately represent the semantics of various event types and physical nodes, respectively. For the proposed topological temporal causal transformer model, short-term and long-term dependencies in the sequence can be captured. Specifically, three separable transformer model variants, namely the topological transformer model, the causal transformer model, and the temporal transformer model, are stacked in sequence to jointly model the spatial and temporal correlations between event types.

[0187] refer to Figure 6 This embodiment further discloses a fault location system for a telecommunication network, including:

[0188] An initialization module, used to initialize the topological temporal causal graph based on the topological structure and time steps of the event sequence;

[0189] A joint operation module, used for combining the convolution operations in the topological domain and the time domain into a joint convolution operation in the topological time domain, and generating a topological time causal graph;

[0190] The information injection module is used to define adaptive embedding based on GCN (graph convolutional network) to inject the initial causal structure information of event type and physical node topology into the event sequence embedding;

[0191] Model building module, used to build topological temporal causal transformer models and capture global dependencies; also used to,

[0192] Optimizing the topological temporal causal transformer model parameters by maximizing the log-likelihood function; and also for,

[0193] constructing a continuous-time conditional intensity function based on the output of the topological-temporal causal transformer model;

[0194] The log-likelihood maximization function is constructed through the continuous-time conditional intensity function; it is also used for

[0195] The causal weights are updated iteratively and finally saved as a causal graph between event types to locate the fault.

[0196] This embodiment further discloses a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned telecommunication network fault location method is implemented.

[0197] This embodiment further discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned telecommunication network fault location method when executing the program.

[0198] This embodiment further discloses a computer program product, including a computer program, which implements the above-mentioned telecommunication network fault location method when executed by a processor.

[0199] A person skilled in the art can understand that all or part of the steps for implementing the above-mentioned method implementation method can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium, which, when executed, executes the steps of the above-mentioned method implementation method; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, etc., various storage media that can store program codes.

[0200] The device implementation described above is merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present implementation scheme. Those of ordinary skill in the art may understand and implement it without creative effort.

[0201] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each implementation method or some parts of the implementation method.

[0202] The specific manner of executing the operations of each unit in the above-mentioned implementation mode has been described in detail in the implementation mode of the method, and will not be elaborated here.

[0203] The above are only preferred implementations of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for locating a fault in a telecommunication network, characterized in that: The method comprises: Initialize the topological temporal causal graph based on the topological structure and time steps of the event sequence; The convolution operations in the topological domain and the time domain are combined into a joint convolution operation in the topological time domain, and a topological time causal graph is generated; Define adaptive embedding based on GCN (graph convolutional network) to inject the initial causal structure information of event type and physical node topology into the event sequence embedding; Build a topological temporal causal transformer model and capture global dependencies; Optimizing the topological temporal causal transformer model parameters by maximizing a log-likelihood function; The causal weights are updated iteratively and finally saved as a causal graph between event types to locate the fault.

2. The method for locating a fault in a telecommunication network according to claim 1, wherein: The method of combining the convolution operations in the topological domain and the time domain into a joint convolution operation in the topological time domain and generating a topological time causal graph includes: Let B∈R N×N is a weighted causal matrix. Given an event type n,n′∈N, B n,n′ Characterize the causal strength from type n to type n′, let (n,v,t) and (n′,v′,t′) be the events observed in the event sequence E, where Calculate B n,n′ The topological time causal graph is convoluted and defined as follows: Among them, φ (v,t),(v′,t′) is the convolution kernel of the topological domain, is the convolution kernel of the time domain, n, n′ is the given event type, G V represents the topological structure in the network, and T represents the dimension of time.

3. The method for locating a fault in a telecommunication network according to claim 1, wherein: The method of defining adaptive embedding based on GCN (graph convolutional network) and injecting initial causal structure information of event type and physical node topology into event sequence embedding includes: Build an embedding layer based on GCN and inject the topological, causal, and temporal information of event observations; Adaptive encoding based on GCN, specifically: Physical node encoding, let Represents the vertex subsequence of the alarm event, initializes the physical node embedding matrix D V ∈R M×|V| , where D V The jth column represents the vertex v j M-dimensional embedding, assuming is the input of the kth layer of GCN, where At this time, the topological graph convolution is defined as inputting the topological relationship of the physical nodes into the node embedding. At this time, the topological graph convolution is: Where W V is the parameter matrix, matrix A is the fixed topology structure, is the graph structure learned dynamically by parameterizing A; Modify the embedding of the event vertex subsequence to: in is the one-hot encoding matrix of the event vertex subsequence, is the vertex v j A one-hot encoding of , j∈V; Event type code, let Represents the type subsequence of the alarm event. The event types in N are mapped to high-dimensional features using a trainable linear projection. The embedding matrix of the event type is represented as D N ∈R M×|N| , where M is the embedding dimension and the matrix D N The jth column corresponds to event type n j The M-dimensional embedding matrix D N Convert discrete event types into continuous real-valued vectors; Construct a causal graph convolution to fuse the causal semantic information into the event type embedding matrix to obtain the feature representation of event types at the causal level. represents the input of the kth layer of GCN, where Define semantic graph convolution as capturing patterns in causal graphs: in is the parameterized representation of the trainable matrix B, σ is the activation function, W N The parameter matrix semantic graph convolution is defined to capture the intrinsic pattern of the causal graph; after the training is stable, Infer the most appropriate causal graph structure for the event type and reshape the embedding of the event type subsequence into: in is the one-hot encoding matrix of the event type subsequence, the vector Is the event type n j A one-hot encoding of , j∈N; Time code, order Initialize the learnable time embedding matrix D for the time subsequence of the alarm event T ∈R M×|T| , where D T The jth column of represents the time step t j The M-dimensional embedding of the time subsequence E of the alarm event T is encoded into an M×|T|-dimensional matrix X t′ .

4. The method for locating a fault in a telecommunication network according to claim 1, wherein: The method of constructing a topological temporal causal transformer model and capturing global dependencies includes: A topological transformer model is constructed, including a multi-head self-attention layer (MHSA) and a fully connected feedforward layer (FFN). The node embedding features are input into the topological transformer, and the multi-head self-attention layer (MHSA) and the fully connected feedforward layer (FFN) are used to sequentially integrate the information of all alarm times. Constructing a causal transformer model, fusing the output of the topological transformer model and event type embedding to encode high-dimensional features for each event; A temporal transformer model is constructed to encode the temporal information of alarms and capture the global contextual relations.

5. The method for locating a fault in a telecommunication network as claimed in claim 4, characterized in that: After constructing the topological temporal causal transformer model and capturing global dependencies, the method further includes: constructing a continuous-time conditional intensity function based on the output of the topological-temporal causal transformer model; The log-likelihood maximization function is constructed through the continuous-time conditional intensity function.

6. A fault location system for a telecommunication network, characterized in that: include: An initialization module, used to initialize the topological temporal causal graph based on the topological structure and time steps of the event sequence; A joint operation module, used for combining the convolution operations in the topological domain and the time domain into a joint convolution operation in the topological time domain, and generating a topological time causal graph; The information injection module is used to define adaptive embedding based on GCN (graph convolutional network) to inject the initial causal structure information of event type and physical node topology into the event sequence embedding; A model building module to build a topological temporal causal transformer model and capture global dependencies; Also used for The topological temporal causal transformer model parameters are optimized by maximizing the log-likelihood function; and further used for, The causal weights are updated iteratively and finally saved as a causal graph between event types to locate the fault.

7. The fault location system for a telecommunication network as claimed in claim 6, characterized in that: The model building module is also used to: constructing a continuous-time conditional intensity function based on the output of the topological-temporal causal transformer model; The log-likelihood maximization function is constructed through the continuous-time conditional intensity function.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 5 is implemented.

10. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the method according to any one of claims 1 to 5.

Images