Search

CN-122001468-A - Optical communication network fault diagnosis method based on multi-source data fusion

CN122001468ACN 122001468 ACN122001468 ACN 122001468ACN-122001468-A

Abstract

The invention discloses an optical communication network fault diagnosis method based on multi-source data fusion, which comprises the following steps of collecting multi-source data to generate an input tensor, configuring a segment-level circulation memory slot for a segment sequence, inputting a window sequence into a two-way gating circulation unit of an improved type TranAD to obtain local time sequence representation, splicing the state of the segment-level circulation memory slot and the local time sequence representation according to a preset position index, inputting the segment into a multi-layer transducer encoder of the improved type TranAD to encode, setting a memory consistency bridge, and outputting a diagnosis result by inputting the multi-layer encoding representation into a double-flow decoding structure of the improved type TranAD to obtain an abnormal score. The invention is suitable for the intelligent fault diagnosis scene driven by multi-source data in the optical communication network, and has the advantages of high diagnosis precision, high response speed, strong cross-level association and good expandability.

Inventors

  • GENG WENHAI
  • Shou Yesheng

Assignees

  • 北京黎阳之光科技有限公司

Dates

Publication Date
20260508
Application Date
20260319

Claims (9)

  1. 1. The optical communication network fault diagnosis method based on multi-source data fusion is characterized by comprising the following steps: Firstly, collecting multi-source data, preprocessing the multi-source data, and generating an input tensor; Step two, segment division is carried out on the input tensor to obtain a segment sequence arranged in time sequence, a segment-level circulating memory slot is configured for the segment sequence, and single reading and writing are carried out at the boundary of the adjacent segment sequence; Inputting the window sequence in each segment into a bidirectional gating circulation unit of an improved TranAD to obtain a local time sequence representation of the current segment; Splicing the state of the segment-level circulating memory slot of the previous segment as an additional sequence unit and the local time sequence representation of the current segment according to a preset position index, and inputting the splicing result as a segment with enhanced memory; Step five, inputting the segments into a multi-layer transducer encoder of the improved TranAD for encoding to obtain multi-layer encoding representation, setting a memory consistency bridge, mapping the multi-layer encoding representation to a unified memory semantic space, and updating the segment-level circulating memory slot state at the segment sequence boundary according to the mapping result; And step six, inputting the multi-layer coding representation into a double-stream decoding structure of the improved TranAD, obtaining an abnormal score based on the reconstruction result and the prediction result generated by the reconstruction solution wharf and the prediction solution wharf, generating a fault candidate unit set according to the abnormal score and the topology record of the optical communication network, carrying out structured summarization on the fault candidate unit set, and outputting a diagnosis result.
  2. 2. The method for diagnosing a fault in an optical communication network based on multi-source data fusion according to claim 1, wherein the first step specifically comprises: The multi-source data comprises optical layer measurement data, network layer operation data, equipment log data and environment monitoring data; the pretreatment is specifically as follows: performing time alignment on the multi-source data according to the unified time reference to construct a multi-dimensional time sequence data set; and combining the multi-dimensional time sequence data set into a multi-source feature vector with set dimensions according to a preset field sequence, and windowing the time sequence by adopting the set window length and the set sliding step length to generate an input tensor.
  3. 3. The method for diagnosing a fault in an optical communication network based on multi-source data fusion according to claim 1, wherein the step two specifically comprises: Dividing an input tensor into a continuous segment sequence according to a time sequence, wherein each segment corresponds to a window sequence, the window sequence consists of N windows generated according to a set window length and a set sliding step length, and the windows keep an increasing sequence on a time index; Establishing a segment index table for a segment sequence, and recording a start time stamp, an end time stamp, a start window sequence number and an end window sequence number of each segment; Configuring a segment-level circulating memory slot for a segment sequence, and initializing a segment-level circulating memory slot state, wherein the segment-level circulating memory slot consists of M memory slot units and has a set slot number and a set feature dimension; setting a single read-write flow at the boundary of the adjacent segment sequence, executing one time of reading the segment-level circulating memory slot state of the previous segment when the current segment starts, and executing one time of writing the segment-level circulating memory slot state of the current segment when the current segment ends; during the processing in the section, the state of the section-level circulating memory slot is kept unchanged, and the time position of the window sequence is not rearranged; establishing position mapping for the segment-level circulating memory slot and the segment sequence, and setting memory slot units at the head of the window sequence when splicing; outputting the segment sequence and the state of the corresponding segment-level circulation memory slot according to the sequence of the segment index table.
  4. 4. The method for diagnosing a fault in an optical communication network based on multi-source data fusion according to claim 1, wherein said step three specifically comprises: Sequentially selecting a current segment from the segment sequences, and reading a window sequence corresponding to the current segment; Inputting the improved TranAD bidirectional gating circulation unit according to the sequence of the window sequence in the segment index table, wherein the sequence is increased according to the time index; The bidirectional gating circulation unit comprises a forward gating circulation sub-network and a backward gating circulation sub-network; The forward gating circulation sub-network receives window sequence input in the time index increment direction and generates a forward hidden state sequence; The backward gating circulation sub-network receives the input of the same window sequence according to the decreasing direction of the time index and generates a backward hidden state sequence; Forming a time step representation for the forward hidden state and the backward hidden state of each time step according to a set combination mode; Generating a local time sequence representation of the current section according to the time step representations in a set aggregation mode; Setting filling and masking processing is adopted for the window sequence with insufficient length, and filling bits do not participate in state updating of the bidirectional gating circulating unit; The time position of the window sequence is not rearranged in the processing process, and the local time sequence representation of the current segment is output together with the current segment identification.
  5. 5. The method for diagnosing a fault in an optical communication network based on multi-source data fusion according to claim 1, wherein the fourth step specifically comprises: according to the position mapping established for the segment-level circulating memory slot and the segment sequence, the memory slot unit is arranged at the head of the window sequence when being spliced with the window sequence of the current segment; sequentially splicing the segment-level circulating memory slot state and the local time sequence representation of the current segment according to the set position index, wherein the index is kept to be increased in an increasing way and the cross-displacement rearrangement is not carried out in the splicing process; When the state of the segment-level circulating memory slot is inconsistent with the characteristic dimension of the local time sequence representation, adopting a set dimension pair Ji Yingshe to adjust the state of the segment-level circulating memory slot and the local time sequence representation to be consistent with the dimension, and then executing splicing; recording the current segment identification and window index information in the spliced data; and inputting and outputting the spliced result as a memory-enhanced segment.
  6. 6. The method for diagnosing a fault in an optical communication network based on multi-source data fusion according to claim 1, wherein the fifth step specifically comprises: Inputting the memory-enhanced segment into a modified TranAD multi-layer transducer encoder, the multi-layer transducer encoder being composed of sequentially stacked encoding layers, each encoding layer sequentially including a self-attention sub-layer, a feed-forward sub-layer, a residual connection and layer normalization; maintaining the sequence index of the memory-enhanced segment input unchanged during the processing and adopting a set mask for the filling position; outputting corresponding hierarchical coding representations according to the stacking sequence of the coding layers and forming a multi-layer coding representation; setting a memory consistency bridge, and mapping hierarchical coding representations of all coding layers to a unified memory semantic space to obtain a mapping result for memory write-back; Updating the segment-level circulating memory slot state at the segment sequence boundary according to the mapping result, and updating according to the set write-back order and write-back granularity; a set mapping and writing-back rule is adopted between each coding layer and the segment-level circulation memory slot, and the mapping and writing-back rule is kept unchanged in the same segment; the output contains the multi-layer encoded representation and updated state of the segment-level circular memory slot.
  7. 7. The method for diagnosing a fault in an optical communication network based on multi-source data fusion according to claim 6, wherein said setting up a memory consistency bridge performs mapping to a unified memory semantic space on the hierarchical coding representation of each coding layer to obtain a mapping result for memory write-back, and specifically comprises: A layer side mapping unit of a memory consistency bridge is input for the hierarchical coding representation of each coding layer, the hierarchical coding representation is adjusted to the target dimension of the unified memory semantic space according to the set dimension alignment mapping, and the intermediate alignment representation is output; performing effective position screening on the middle alignment representation according to mask settings of the time index and the segment sequence boundary; The mask setting of the segment sequence boundary is a segment boundary mask matrix generated according to the starting window position and the ending window position in the segment index table; Reserving intermediate alignment representations corresponding to the positions with valid values in the segment boundary mask matrix, removing the intermediate alignment representations corresponding to the positions with invalid values, and outputting the effective alignment representations after screening according to the time index sequence; Aggregating the effective alignment representation according to a set aggregation rule to obtain a segment level alignment representation of the current coding layer; Merging the segment level alignment representations of the coding layers according to the stacking sequence of the coding layers and a set merging rule to generate segment level merging alignment representations; And according to the position mapping established for the segment-level circulating memory slot and the segment sequence, dividing the segment-level merging alignment representation into write-back entries corresponding to the number of memory slot units according to the set write-back granularity, and forming a mapping result for memorizing the write-back.
  8. 8. The method for diagnosing a fault in an optical communication network based on multi-source data fusion according to claim 1, wherein the sixth step specifically comprises: the double-stream decoding structure comprises a reconstruction solution wharf and a prediction solution wharf; The reconstruction decoding head outputs a reconstruction sequence corresponding to the current window sequence one by one, and the prediction decoding head outputs a prediction sequence aligned to the time index; generating an abnormal scoring sequence at a window level for the multi-layer coding representation, the reconstruction sequence and the prediction sequence according to the set error metric and the combination rule, and recording a corresponding time index for each scoring; outputting an abnormal scoring sequence according to the time index sequence; Calling an optical communication network topology record, and completing mapping from the abnormal time index to the port, the link and the equipment according to the corresponding relation between the time index in the abnormal scoring sequence and the port identification, the link identification and the equipment identification in the optical communication network topology record to obtain a fault candidate unit set; And carrying out structured summarization on the fault candidate unit set according to a set merging and deduplication rule, generating and outputting a diagnosis result containing fault unit identification, occurrence time, type identification and confidence coefficient value.
  9. 9. The method for diagnosing a fault in an optical communication network based on multi-source data fusion according to claim 8, wherein the reconstructing decoding head outputs a reconstructed sequence corresponding to the current window sequence one by one, and the predicting decoding head outputs a predicted sequence aligned to the time index, specifically comprising: a reconstruction decoding head receives the multi-layer coded representation, and decodes the multi-layer coded representation by adopting a bidirectional attention structure; reading multi-layer coding representation in the time index range of the current section and generating a query vector, a key vector and a value vector for each time step; Canceling the setting of the causal mask, and calculating the correlation coefficient of each time step query vector and all time step key vectors according to a time index full-connection mode; Performing weighted summarization on the value vectors of the corresponding time steps according to the correlation coefficients to obtain decoding representation of each time step; The decoding representation of each time step is sequentially processed by the self-attention sub-layer, the feedforward sub-layer, the residual connection and the layer normalization and output to the next decoding layer; performing linear mapping on the decoded representation at the last layer to restore to the input characteristic dimension, and generating a reconstruction sequence corresponding to the current section window one by one according to the sequence order of the window; The prediction solution terminal receives the same multi-layer coding representation, and adopts a causal attention structure to conduct sequence prediction; reading multi-layer coding representation in the time index range of the current section, and generating a query vector, a key vector and a value vector for each time step; Setting a lower triangular mask matrix to shield information of future time steps, so that the query vector of each time step is only calculated in a correlated manner with the query vector and key vectors of the historical time steps; Performing weighted summation on the value vectors of the corresponding time steps according to the correlation coefficients to generate decoding representations which are arranged in ascending order according to the time indexes; The decoding representation is sequentially subjected to self-attention sub-layer, feedforward sub-layer, residual connection and layer normalization processing and output to the next decoding layer; And performing linear mapping on the decoded representation at the last layer to restore to the input characteristic dimension, and generating a prediction sequence corresponding to the time steps one by one according to the time index sequence.

Description

Optical communication network fault diagnosis method based on multi-source data fusion Technical Field The invention relates to the technical field of communication network operation and maintenance, in particular to an optical communication network fault diagnosis method based on multi-source data fusion. Background In the context of rapid development of optical communication networks, the scale and structure of the networks are increasingly complex, the number of devices is continuously increased, and the link density is continuously increased, so that the operation and maintenance of the optical communication system face great challenges. In order to ensure the stability and reliability of the network, how to efficiently and accurately realize the network fault diagnosis has become a core problem in operation management. The existing fault detection method of the optical communication network is mostly dependent on information of a single data source, for example, analysis is only carried out based on optical layer measurement data or network layer operation state, the accuracy of the method is low when the method faces to cross-layer faults or multi-source interference factors, and heterogeneous fault characteristic expression in a complex environment is difficult to deal with. With the deep application of artificial intelligence technology, the anomaly detection and fault diagnosis method based on deep learning is gradually applied to the operation and maintenance of a communication network. Among them, time series modeling is widely adopted, such as modeling and anomaly localization of data using a gated loop unit (GRU), a long-short-term memory network (LSTM), or a self-attention mechanism (transducer). These approaches improve the sensitivity and automation level of fault detection to some extent. However, most models are still limited to processing single pieces of continuous data, lack of retention of long-term context, and difficulty in state transfer and memory interaction across time periods, resulting in failure of diagnostic information to fully cover latent anomalies. Meanwhile, the lack of a structured mapping and coordination mechanism between the coded representation and the memory state affects the reasoning logic consistency of the model. The fusion capability of the prior art on multi-source data is limited, heterogeneous information sources such as an optical layer, a network layer, equipment logs and environment monitoring are difficult to integrate effectively, and the prior art has defects in time alignment, feature combination and context modeling. In the aspect of model structural design, deep exploration of a segment-level cyclic memory mechanism is not available, state information of a history segment cannot be fully utilized to assist generation of a current time sequence representation, and fault identification with characteristics such as slowly-varying property and burstiness in an optical communication network is not sensitive enough. Therefore, how to provide an optical communication network fault diagnosis method based on multi-source data fusion is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide an optical communication network fault diagnosis method based on multi-source data fusion, which fully utilizes deep time sequence analysis technologies such as multi-source time sequence fusion, an improved TranAD model, a two-way gating circulation unit (BiGRU) structure, a segment-level circulation memory slot mechanism, a memory consistency bridge and the like to construct an intelligent diagnosis framework capable of maintaining state memory and mapping active characteristics across time segments. The method realizes unified expression of multi-source heterogeneous information by joint modeling of optical layer measurement data, network layer operation data, equipment logs and environment monitoring data, establishes a consistency updating mechanism of cross-segment context transfer and multi-layer coding memory by introducing a segment-level circulating memory slot and improved TranAD-BiGRU structure, comprehensively analyzes a reconstruction sequence and a prediction sequence by utilizing a double-stream decoding structure, generates window-level abnormal scores, and realizes intelligent identification and output of fault positions, types and confidence. According to the embodiment of the invention, the optical communication network fault diagnosis method based on multi-source data fusion comprises the following steps: Firstly, collecting multi-source data, preprocessing the multi-source data, and generating an input tensor; Step two, segment division is carried out on the input tensor to obtain a segment sequence arranged in time sequence, a segment-level circulating memory slot is configured for the segment sequence, and single reading and writing are carried out at the boundary of the adjacent segment