Optical network data processing method and system based on large language model

Abstract

This invention provides a method and system for optical network data processing based on a large language model, belonging to the field of optical communication network operation and maintenance technology. The method includes: mapping each historical numerical sequence in the optical network time-series data to a feature embedding vector and generating a labeled embedding sequence; constructing text prompt words containing task prompt information; modally aligning the labeled embedding sequence with the text features of the text prompt words to generate an aligned input representation; and inputting the aligned input representation into a fine-tuned pre-trained inference model to obtain a text response associated with the time-series data. This application achieves deep integration of numerical data and text semantics by mapping complete historical numerical sequences to feature vectors and aligning them with text features. It solves the problem of existing numerical models lacking semantic understanding capabilities, and can directly output logically interpretable analytical text, improving the processing efficiency and readability of optical network operation and maintenance data.

Description

Technical Field

[0001] This invention relates to the field of optical communication network operation and maintenance technology, and in particular to an optical network data processing method and system based on a large language model. Background Technology

[0002] As the core of modern communication infrastructure, the reliability of optical networks directly affects the quality of data transmission and the continuity of services. With the increasing scale and complexity of optical networks, equipment operating parameters exhibit multi-dimensional coupling characteristics. Analyzing these data to achieve accurate and timely fault prediction is of great significance for improving network stability and reducing operation and maintenance costs.

[0003] To achieve early identification and prediction of optical network faults, existing technologies typically employ time series analysis methods based on deep learning. For example, algorithms such as Long Short-Term Memory (LSTM) networks or Recurrent Neural Networks (RNNs) are used to construct predictive models. These approaches usually take historical performance metrics collected by the optical network as input, extract and process the temporal characteristics of the data through the model, and ultimately output numerical predictions of future performance metrics, or probability values ​​indicating whether the current network is at risk of failure.

[0004] However, the aforementioned existing technologies have significant limitations in practical applications. Because these models primarily rely on numerical fitting and statistical regularities, they can only output abstract numerical values ​​or probability labels, lacking semantic understanding and logical reasoning capabilities regarding the physical meaning behind the data. They cannot directly generate semantic text containing causal analysis or operational suggestions. This forces maintenance personnel to perform time-consuming manual secondary interpretation and analysis of the model's output numerical results, resulting in low processing efficiency for optical network maintenance data and poor readability of the data processing results. Summary of the Invention

[0005] This invention provides a method and system for optical network data processing based on a large language model, which addresses the shortcomings of existing technologies and improves the processing efficiency and readability of optical network operation and maintenance data.

[0006] This invention provides a method for optical network data processing based on a large language model, comprising the following steps: Each historical numerical sequence in the time-series data of the optical network is mapped to a feature embedding vector, and a labeled embedding sequence is generated based on all the said feature embedding vectors; Construct text prompt words containing task prompt information, and modally align the embedded sequence of the markers with the text features of the text prompt words to generate an aligned input representation; The aligned input representation is fed into a pre-trained inference model to obtain the text response output by the inference model that is associated with the time series data; wherein the inference model is obtained by fine-tuning a large language model.

[0007] According to the optical network data processing method based on a large language model provided by the present invention, the step of modally aligning the token embedding sequence with the text features of the text prompt words to generate an aligned input representation includes: The cross-attention mechanism is used to process the labeled embedding sequence and the text features to generate an aligned input representation; Specifically, for each attention head of the cross-attention mechanism, a query vector is generated based on the labeled embedding sequence, and a key vector and a value vector are generated based on the text features; The output of the current attention head is generated based on the query vector, the key vector, and the value vector. The outputs of all the attention heads are concatenated and then linearly projected to generate the aligned input representation.

[0008] According to the optical network data processing method based on a large language model provided by the present invention, the aligned input representation is calculated by the following formula: ; ; In the formula, K represents the number of attention heads. For the output of the k-th attention head, Let k be the query vector of the attention head. Let k be the key vector of the attention head. Let T be the value vector of the k-th attention head, and let T be the transpose sign. Let be the dimension of the feature vector in the attention head. For normalized exponential functions, This represents the aligned input. For concatenation functions, This is for outputting the projection matrix.

[0009] According to the present invention, a method for processing optical network data based on a large language model, wherein mapping each historical numerical sequence in the time series data of the optical network to a feature embedding vector includes: For each network variable in the time series data, obtain the complete historical value sequence within a preset time window; The historical numerical sequence is processed by a fully connected projection using a multilayer perceptron. The historical numerical sequence is compressed and mapped into a fixed-dimensional feature embedding vector, wherein the feature embedding vector represents the overall dynamic trend of the network variables within the preset time window.

[0010] According to the present invention, an optical network data processing method based on a large language model is provided, wherein the inference model includes multiple inference modules and multiple candidate low-rank adaptation modules; each inference module includes a Transformer layer and a router disposed before the Transformer layer; The step of feeding the aligned input representation into a pre-trained inference model to obtain the text response output by the inference model associated with the time series data includes: For any one of the inference modules, the router calculates the matching score between the input data and each of the candidate low-rank adaptation modules, and based on the matching score, selects the matching target low-rank adaptation module from the plurality of candidate low-rank adaptation modules and activates the target low-rank adaptation module. The input data is collaboratively inferred by the Transformer layer and the target low-rank adaptation module to generate output data. The input data of the first inference module is the aligned input representation; the input data of the subsequent inference module is the output data of the previous inference module; and the output data of the last inference module is the text response.

[0011] According to the present invention, an optical network data processing method based on a large language model is provided, wherein the inference model is obtained through the following fine-tuning steps: Acquire training samples, which include historical fault tickets of optical networks and simulated expert dialogues; the simulated expert dialogues are generated by a general large language model. Keeping the parameters of the Transformer layer frozen, the parameters of the candidate low-rank adaptation module and the router are updated and trained using the training samples to generate the inference model.

[0012] This invention also provides an optical network data processing system based on a large language model, comprising the following modules: The first processing module is used to map each historical numerical sequence in the time series data of the optical network to a feature embedding vector, and generate a labeled embedding sequence based on all the said feature embedding vectors; The second processing module is used to construct text prompt words containing task prompt information, and modally align the embedded sequence of the markers with the text features of the text prompt words to generate an aligned input representation; The third processing module is used to input the aligned input representation into a pre-trained inference model to obtain the text response output by the inference model that is associated with the time series data; wherein the inference model is obtained by fine-tuning a large language model.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the optical network data processing method based on the large language model as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the optical network data processing method based on the large language model as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the optical network data processing method based on any of the above-described large language models.

[0016] In summary, one or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: This invention maps the complete historical numerical sequence of each variable in an optical network to feature embedding vectors and generates a labeled embedding sequence, thus preserving the overall trend of network parameters within a preset time window. This achieves semantic-level compression and representation of multivariate data, thereby improving the model's efficiency in extracting collaborative features between different parameters. Furthermore, by modally aligning the labeled embedding sequence with text prompts containing task information, numerical features are mapped to a semantic space consistent with text features, ensuring that the large language model can accurately understand the physical meaning of numerical data in conjunction with the specific task context. Finally, by inputting the aligned data into a fine-tuned inference model to generate a text response, it directly outputs analytical conclusions or suggestions related to the time series data. This solves the problem that traditional data processing methods can only output numerical indicators without logical explanation, improving the processing efficiency and readability of optical network operation and maintenance data. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the optical network data processing method based on a large language model provided by the present invention.

[0019] Figure 2 This is a flowchart of optical network data processing provided by the present invention.

[0020] Figure 3 The schematic diagram of the multi-parameter fault prediction and LLM fine-tuning is shown.

[0021] Figure 4 This is a schematic diagram illustrating the performance data acquisition process during the preparation of the multimodal dataset provided by this invention.

[0022] Figure 5 This is a flowchart of the fine-tuning dataset construction process in the preparation of multimodal datasets provided by the present invention.

[0023] Figure 6 This is a schematic diagram comparing the improved model provided by this invention with various existing open-source models in terms of training loss convergence characteristics.

[0024] Figure 7 This is a schematic diagram comparing the performance indicators of the improved model provided by this invention with those of various existing open-source models in terms of multi-parameter trend prediction and fault diagnosis.

[0025] Figure 8 This is a radar chart comparing the improved model provided by this invention with various existing open-source models in terms of multi-dimensional indicators of text generation quality and computational efficiency.

[0026] Figure 9 This is a schematic diagram illustrating the performance analysis of the improved model provided by this invention in few-shot learning generalization under different proportions of training data.

[0027] Figure 10 This is a schematic diagram of an interaction example of the optical network active defense framework based on a large language model provided by this invention in a real-world dialogue scenario.

[0028] Figure 11 This is a schematic diagram of the structure of the optical network data processing system based on a large language model provided by the present invention.

[0029] Figure 12 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0031] It should be noted that in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. The terms "upper," "lower," etc., indicating orientation or positional relationships according to the accompanying drawings, are only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0032] The terms "first," "second," etc., used in this invention are used to distinguish similar objects, not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class, without limiting the number of objects; for example, a first object can be one or more. Furthermore, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0033] The following is combined with Figures 1 to 12 This invention describes the optical network data processing method, system, electronic device, storage medium, and computer program product based on a large language model.

[0034] This application provides a method for optical network data processing based on a large language model. Before introducing the embodiments of this application, the execution subject of these embodiments is first described. The method provided in this application is executed by a server, which includes, but is not limited to, a workstation, a network management controller, a cloud computing platform, or a distributed computing system composed of multiple computing nodes. In a specific application scenario, the server is deployed in the network operation and maintenance center of the optical transport network to monitor, predict, and maintain the operating status of the optical network.

[0035] Reference Figure 1 , Figure 1 This is a flowchart illustrating the optical network data processing method based on a large language model provided by the present invention. Figure 1As shown, this embodiment of the invention provides a method for optical network data processing based on a large language model. This method is executed by an intelligent optical network operation and maintenance server and specifically includes the following steps: Step 1: Map each historical numerical sequence in the time series data of the optical network to a feature embedding vector, and generate a labeled embedding sequence based on all feature embedding vectors.

[0036] In complex optical network operation and maintenance scenarios, the optical network intelligent operation and maintenance server first obtains time-series data of the optical network from the network management system or underlying devices of the optical transport network (OTN). This time-series data contains performance indicators describing the operating status of the optical network in multiple dimensions, such as input optical power, output optical power, optical signal-to-noise ratio (OSNR), bit error rate (BER), board temperature, and laser bias current. These data change continuously over time, constituting a multivariate time series.

[0037] For the acquired multivariate time series data, the server does not simply treat it as independent numerical points, but focuses on its overall evolutionary characteristics over time. Specifically, the server identifies each network variable (i.e., each performance indicator channel) in the time series data and extracts the complete data of that network variable within a historical time window, forming a historical numerical sequence. Subsequently, the server performs a mapping operation through its internally integrated sequence encoding module, directly mapping this historical numerical sequence containing multiple time steps into a fixed-length high-dimensional vector, namely the feature embedding vector. This feature embedding vector is not merely a collection of numerical values, but a highly semantically compressed and abstract representation of the overall dynamic trend, fluctuation pattern, and statistical characteristics of the network variable within that time window.

[0038] The server performs the above mapping operation on each network variable in the time series data, thereby obtaining multiple feature embedding vectors. The server arranges and combines these feature embedding vectors in a predetermined order (e.g., variable type order or spatial location order) to generate a sequence of multiple vectors, namely the token embedding sequence. This token embedding sequence completely preserves the spatiotemporal characteristics of the current operating state of the optical network in numerical mode.

[0039] Step 2: Construct text prompts containing task hints, and modally align the token embedding sequence with the text features of the text prompts to generate the aligned input representation.

[0040] To enable the large language model to understand the physical meaning of numerical data and execute specific operational tasks, the server needs to incorporate contextual information from natural language. Based on current operational needs (such as fault prediction, root cause analysis, or maintenance suggestion generation), the server constructs a text prompt containing task hints. This prompt typically includes a description of the optical network topology, device type, current link status, and specific analysis instructions (e.g., "Please analyze the current optical power trend and predict future status"). The server then performs text encoding on this prompt, extracting its semantic information to obtain text features.

[0041] Subsequently, the server performs modality alignment. Since the labeled embedding sequence belongs to the numerical modality, while the text prompts belong to the natural language modality, their distributions differ in the original feature space. Through modality alignment, the server establishes a deep semantic association between the labeled embedding sequence and the text features, mapping or transforming the numerical labeled embedding sequence into a semantic space compatible with the text features. This process ensures that subsequent models, when processing data, can not only see the changes in numerical values ​​but also understand the physical meaning behind these values ​​by combining prior knowledge (such as topological constraints and task objectives) contained in the text prompts. After modality alignment, the numerical features and text features are fused or integrated into a unified input form containing rich contextual information—the aligned input representation.

[0042] Step 3: Input the aligned input representation into the pre-trained inference model to obtain the text response output by the inference model that is associated with the time series data; wherein, the inference model is obtained by fine-tuning the large language model.

[0043] The server uses the aligned input representation generated above as input data and feeds it into the pre-trained inference model. This inference model is a domain-specific model obtained by fine-tuning training data containing optical network domain expertise (such as historical fault tickets, expert experience bases, etc.) based on a general large language model (LLM).

[0044] During the inference process, this inference model leverages its powerful self-attention mechanism and logical reasoning capabilities to deeply analyze the performance parameter variation trends, coupling relationships between parameters, and their correlation with task instructions inherent in the aligned input representation. The model no longer simply outputs a single failure probability or numerical prediction result; instead, based on its learned domain knowledge, it generates a coherent, logical, and human-readable natural language text—a text response. The content of this text response is directly related to the input time-series data, representing a deep analysis of the optical network's operational state.

[0045] The technical solution provided in this invention effectively transforms heterogeneous optical network numerical data into semantic representations that can be understood and processed by a large language model by mapping the complete historical sequence of each variable of the optical network into feature embedding vectors and modally aligning them with text prompts containing task information. This method breaks through the limitation of traditional numerical prediction models that can only output quantitative indicators. It utilizes a fine-tuned inference model to achieve a leap from numerical monitoring to natural language analysis, directly outputting semantically interpretable text responses. This assists maintenance personnel in quickly understanding the network status, improving the intelligence level and decision-making efficiency of optical network operation and maintenance.

[0046] As another implementation method combined with specific business processes, this embodiment will be described in conjunction with the appendix to the specification. Figure 2 This paper provides a comprehensive explanation of the above-mentioned optical network data processing methods based on large language models.

[0047] Figure 2 This is a flowchart of optical network data processing provided by the present invention. Figure 2 As shown, the optical network intelligent operation and maintenance server first initiates the implementation process of the proactive defense recommendations. During the data preparation phase, the server acquires two types of key data in parallel: one type is historical multi-parameter performance data from the optical network's historical operation records (denoted as...). X 1... Xn The first type of data contains time series data covering key indicators such as optical power and bit error rate; the second type contains domain knowledge prompts, which include operational knowledge from optical network experts and specific instructions for this task. These two types of heterogeneous data are sent to the "Time Series Tagging and Modality Alignment Module" for centralized processing. Within this module, the server first performs time series feature extraction, and... X 1... Xn The complete historical sequence of each parameter is mapped to a high-dimensional feature embedding vector, thereby generating a label embedding sequence that preserves the temporal dynamics. Then, the server modally aligns this label embedding sequence with the text features of domain knowledge prompts to eliminate the semantic gap between numerical data and text data and generate a semantically fused input representation.

[0048] The server then inputs the aligned data into the "fine-tuned domain LLM". For example... Figure 2 As shown, this inference model integrates a Dynamic Low-Rank Adaptation (Dynamic LoRA) mechanism, which can dynamically adjust model parameters based on the input optical network state to adapt to different fault scenarios. The inference model first extrapolates the future operating state of the optical network based on the input data, outputting multi-parameter future prediction values ​​(denoted as...). Y 1... YnAt this point, the server doesn't stop at simply outputting numerical values; instead, it uses these predicted values ​​to make intelligent judgments. When the predicted values ​​are within the normal range, the model generates and outputs a "normal report," informing operations personnel that the network is running healthily. However, when the predicted values ​​show potential anomalies (i.e., "predicted value failure"), the model further triggers a deep inference chain, combining domain knowledge to determine the root cause of the failure and generating a "proactive defense operation guide" containing specific operational steps. Both the normal report and the defense recommendations containing the root cause and guidelines are text responses output by the model that are associated with time-series data.

[0049] In a preferred embodiment, the optical network intelligent operation and maintenance server performs modal alignment of the token embedding sequence with the text features of the text prompt words, specifically including the following steps: Step 21: Use the cross-attention mechanism to process the labeled embedding sequence and text features to generate an aligned input representation.

[0050] Because time-series data from optical networks is numerical, while large language models excel at handling textual semantic modalities, there are significant differences in feature distribution between the two. To eliminate this heterogeneity, the server introduces a modality alignment module, whose core architecture employs a multi-head cross-attention (MHCA) mechanism. The server uses this mechanism to establish a deep interaction channel between time-series data and textual prompts, enabling numerical features to "pay attention to" and integrate the semantic context within the text.

[0051] Step 22: For each attention head of the cross-attention mechanism, generate a query vector based on the label embedding sequence, and generate a key vector and a value vector based on the text features.

[0052] Internally, the multi-head cross-attention mechanism uses multiple attention heads in parallel to capture feature associations in different subspaces. For any given attention head, the server executes the following specific role assignment strategy: The server uses the labeled embedding sequence (i.e., feature vectors derived from the historical numerical sequence of the optical network) as the "query" of information. Specifically, the server performs a linear transformation on the labeled embedding sequence using the query weight matrix corresponding to the attention head, generating a query vector (Query, Q). This means that the model will use the real-time performance state of the optical network as a starting point to find relevant explanations or background.

[0053] Simultaneously, the server uses the textual features of the text prompts (i.e., semantic vectors containing information such as topology and instructions) as the "index" and "content" of the information. Specifically, the server performs linear transformations on the textual features using the key weight matrix and value weight matrix corresponding to the attention head, generating key vectors (Key, K) and value vectors (Value, V). This means that the prior knowledge and task instructions in the text prompts constitute the knowledge base being queried.

[0054] Step 23: Generate the output of the current attention head based on the query vector, key vector, and value vector.

[0055] After generating Q, K, and V, the server calculates the relevance score (i.e., attention weight) between the query vector and the key vector. This process measures which semantic information in the text prompt is most relevant to the optical network's numerical features at the current time step. Subsequently, the server uses this relevance score to perform a weighted summation of the value vectors, thereby injecting the semantic information from the text features that best matches the current numerical state into the numerical features, generating the output vector of the current attention head.

[0056] Step 24: After concatenating the outputs of all attention points, perform linear projection to generate the aligned input representation.

[0057] The server collects the output vectors of all parallel computation attention heads and performs a concatenation operation to aggregate information scattered across different feature subspaces. Subsequently, the server uses a linear projection layer to perform feature fusion and dimensionality transformation on the concatenated result, ultimately generating an aligned input representation. This input representation not only preserves the temporal dynamics of the optical network but also deeply integrates the semantic understanding of the text prompts, placing it within the semantic space that the backbone network of a large language model can directly understand and process.

[0058] The technical solution provided in this invention employs a cross-attention mechanism and specifically uses time-series features as queries and text features as keys and values ​​to achieve a clearly targeted modal alignment. This approach forces the model to retrieve background knowledge from text prompts based on the actual numerical fluctuations of the optical network, thereby significantly enhancing the model's ability to understand the implicit semantics within the performance context of the optical network and providing a high-quality input foundation for generating accurate fault diagnoses.

[0059] In one possible implementation, refer to Figure 3 , Figure 3 The schematic diagram of multi-parameter fault prediction and LLM fine-tuning is shown. Figure 3The "Modality Alignment" module in the middle section details the specific hardware / logic implementation path of the solution in the above embodiments. Based on this architecture, the optical network intelligent operation and maintenance server performs the following steps: like Figure 3 As shown, the data is processed into two parallel input streams before entering the modal alignment module.

[0060] First input stream ( Figure 3 The branch above contains time-series data of the optical network. This data is processed by a multilayer perceptron (MLP) and a normalization (norm) layer to generate a labeled embedding sequence. During the modality alignment stage, the server designates this labeled embedding sequence as the source of a query (Q). This means that real-time performance fluctuations of the optical network (such as fluctuations in the optical power curve) are used as "questioners" to actively retrieve relevant information.

[0061] Second input stream ( Figure 3 The branch below contains a text prompt. This prompt includes task instructions, such as "based on previous..." <n>Step size information predicts the next step <t>(Information on steps) and statistical information (such as "the minimum value of the input feature is...").<min_val> The maximum value is<max_val> The median is<median_val> The overall trend is "<rising or falling>...". After being processed by the projection layer, this text prompt generates text features. During the modal alignment stage, the server designates these text features as the source of the key (K) and value (V). This means that the pre-defined expert knowledge and task context are used as the "knowledge base being queried".

[0062] Next, the modal alignment step is performed, namely... Figure 3 The "Modality Alignment" shown here. Specifically, the server utilizes... Figure 3 The "Multi-Head Attention" module performs the core modality alignment calculation. Specifically, for each attention head in this module, the server performs the following operations: Linear transformation generates vectors: The server uses the unique weight matrix of this attention head to perform a linear mapping on the input label embedding sequence to generate a query vector Q; at the same time, it performs a linear mapping on the input text features to generate a key vector K and a value vector V respectively.

[0063] The attention score and weighted server calculates the dot product of the query vector Q and the key vector K. This dot product represents the degree of matching between "the current state of the optical network" and "a certain description in the text prompt" (i.e., the attention score). For example, if the Q vector represents "a sharp drop in optical power", it may produce a high matching score with the text feature in the K vector that represents "risk of fiber breakage".

[0064] Context injection: The server uses the attention score to perform a weighted summation on the value vector V. Through this process, semantic information highly relevant to the current fault features in the text prompt (such as "check connector") is extracted and injected into the current time-series features, generating the output vector of the attention head.

[0065] Furthermore, such as Figure 3 As shown by the output arrow of the "Multi-Head Attention" module, the server concatenates the output vectors obtained from the parallel computation of all attention heads. The concatenated vector contains rich information from different semantic subspaces. Subsequently, the data stream enters... Figure 3 The linear projection layer shown (i.e. Figure 3 The part pointed to by the arrow after the "Multi-Head Attention" module (the server uses this linear projection layer to perform dimensionality transformation and feature fusion on the concatenated vector, ultimately generating the aligned input representation). Figure 3 The text is labeled "TS Tokens Aligned". These "Aligned TS Tokens" are numerical features that incorporate textual semantics. They are then fed into a pre-trained inference model, which is a LoRA-tuned version of the pre-trained Llama-8B model. Finally, through output projection, a text response containing fault prediction and proactive defense suggestions is generated.

[0066] In a preferred embodiment, this embodiment provides a detailed explanation of the specific mathematical calculation process for modal alignment.

[0067] In combination with the above Figure 3 In addition to the cross-attention processing described in the above embodiments, the optical network intelligent operation and maintenance server employs rigorous mathematical operations to ensure the accurate fusion of numerical features and textual features when performing modal alignment. Specifically, the aligned input representation is calculated through the following steps and formulas: First, the server independently computes the output vector for each attention head in the cross-attention mechanism (indexed as k, where k ranges from 1 to K, and K represents the total number of attention heads). The specific calculation follows the standard form of Scaled Dot-Product Attention, as shown in the following formula: ; In this formula: The query vector representing the k-th attention head is obtained by linear transformation of the labeled embedding sequence of the optical network and carries the dynamic features of the time series.

[0068] The key vector representing the k-th attention head is derived from the text features of the text prompt words and is used to provide semantic indexing.

[0069] The value vector representing the k-th attention head also originates from text features and carries specific semantic content.

[0070] In this context, T represents the matrix transpose operation, which is used to adjust the dimensions for dot product operations.

[0071] This represents the dot product of the query vector and the key vector, used to calculate the original attention score.

[0072] This indicates the dimension of the feature vector in the attention head. This is to serve a scaling function, preventing the dot product result from becoming too large and causing subsequent gradient vanishing problems, thus ensuring numerical stability.

[0073] This represents the normalized exponential function, which transforms the scaled dot product into attention weights in the form of a probability distribution. These weights reflect the importance of each part of the text features to the current query.

[0074] Finally, by combining the calculated attention weights with the value vector Multiply (weighted sum) to obtain the output of the k-th attention head. .

[0075] After completing the parallel computation of all K attention heads, the server performs feature aggregation and projection operations, as shown in the following formula: ; In this formula: This represents the concatenation function; the server will... arrive These K output vectors are concatenated along the feature dimension to form a long vector containing information from multiple subspaces.

[0076] This represents the output projection matrix, which is a learnable linear transformation parameter matrix.

[0077] By concatenating the vector with By multiplying, the server fuses the multi-head features and maps them back to the target dimension, ultimately obtaining the aligned input representation. , that is Figure 3 The "Aligned TS Tokens" shown.

[0078] The technical solution provided in this invention clarifies the internal computational logic of cross-attention through specific mathematical formulas. By utilizing scaled dot product and Softmax normalization, the correlation between time-series features and textual semantic features is accurately quantified. Furthermore, through multi-head concatenation and linear projection, effective integration of information from different semantic subspaces is achieved. This precise computational method ensures that the generated input representation retains the key information of the original data to the greatest extent possible, while also adapting to the input requirements of subsequent large language models, providing a solid mathematical foundation for improving the model's inference accuracy.

[0079] like Figure 3 The diagram shows the "time series data" to "Patch embedding" and "MLP" processing sections. In step 1, the optical network intelligent operation and maintenance server performs the operation of mapping each historical numerical sequence to a feature embedding vector. This process does not use the traditional time series slicing (Patching) method, but instead adopts a strategy based on full sequence mapping. The specific steps are as follows: Step 11: For each network variable in the time series data, obtain the complete historical value sequence within the preset time window.

[0080] Time-series data from optical networks typically contains multiple network variables (or channels), such as X1 representing optical power and X2 representing bit error rate. For each variable Xi, the server extracts a segment of continuous data from the data stream within a preset time window (e.g., the past 24 hours or the past 7 days). The key here is that the server obtains the complete sequence data "Xi,:" within that window, where ":" is the slicing symbol, indicating that all elements in that dimension are taken, rather than splitting it into multiple short segments. This operation ensures that the evolution of the variable over time is treated as a single unit.

[0081] Step 12: Perform fully connected projection processing on the historical numerical sequence using a multilayer perceptron.

[0082] The server inputs the complete historical numerical sequence of a specific network variable into a specially designed neural network module. This module preferably employs a Multi-Layer Perceptron (MLP) structure. Unlike Recurrent Neural Networks (RNNs), which process data step-by-step, the MLP receives all time-step data of the entire sequence as input at once through fully connected layers. Internally, the MLP performs high-dimensional feature extraction and projection on the input sequence through weight matrix multiplication and nonlinear transformation of the activation function.

[0083] Step 13: Compress the historical numerical sequence and map it into a fixed-dimensional feature embedding vector, where the feature embedding vector represents the overall dynamic trend of the network variables within a preset time window.

[0084] After processing by the MLP, the original numerical sequence, which may have varied or been quite long, is "compressed" and mapped into a vector with fixed dimensions. That is, the feature embedding vector (also known as the time series token in some contexts).

[0085] This feature embedding vector has a unique physical meaning: it is not a representation of a certain point in time or a certain local waveform in the sequence, but a highly condensed representation of the overall dynamic trend of the network variable within the entire preset time window.

[0086] The server repeats the above process for each network variable, ultimately obtaining a set of feature embedding vectors. Each vector is an independent semantic unit, representing the global state of different performance metrics.

[0087] The technical solution provided in this invention innovatively utilizes MLP to map the complete historical sequence of each variable into a single feature embedding vector, thereby avoiding the temporal fragmentation problem that may be caused by traditional slicing methods and greatly preserving the global evolution characteristics of variables over long time spans. Simultaneously, compressing each variable into a token not only significantly reduces the length of the input sequence for subsequent models and improves computational efficiency, but also enables large language models to capture the relationships between different performance indicators on a variable-by-variable basis, thereby improving the accuracy of fault prediction.

[0088] In a preferred embodiment, this embodiment provides a model architecture for an inference model. Specifically, the inference model includes multiple inference modules and multiple candidate low-rank adaptation modules; each inference module includes a Transformer layer and a router positioned before the Transformer layer.

[0089] Specifically, the main structure of the inference model consists of multiple stacked inference modules (for example, an Llama3 model may contain dozens of inference modules). To achieve efficient and flexible fine-tuning, the server introduces multiple pre-trained candidate low-rank adaptation modules into the model. These candidate low-rank adaptation modules are lightweight parameter matrices (typically containing low-rank decomposed matrices A and B) specifically designed to capture specific types of optical network fault features.

[0090] In addition, each inference module contains a core Transformer layer (containing a self-attention mechanism and a feedforward network) and an intelligent router set up before the Transformer layer.

[0091] Step 3, the process of the server inputting data into the inference model for processing, is essentially a process of data being passed and processed layer by layer among these stacked inference modules. For any one of these inference modules, the data processing flow includes the following steps: Step 31: Calculate the matching score between the input data and each candidate low-rank adaptation module through the router, and select the matching target low-rank adaptation module from multiple candidate low-rank adaptation modules according to the matching score, and activate the target low-rank adaptation module.

[0092] When input data arrives at the current inference module, it first passes through a router. For the first inference module, the input data is the aligned input representation; for subsequent inference modules, the input data is the output data of the previous inference module. The router is a small neural network component used to analyze the feature distribution of the current input data, for example, to identify whether the current data features are biased towards "sudden changes in optical power" or "slow degradation of signal-to-noise ratio".

[0093] Based on the analysis results, the router calculates the matching score (or activation probability) between the current input data and multiple candidate low-rank adaptation modules pre-stored in the model library. This score reflects which LoRA module is best suited to handle the current type of data features.

[0094] The server then executes a selection strategy based on these matching scores (e.g., selecting the top-k modules with the highest scores) to determine one or more of the best-matching modules as target low-rank adaptable modules and activates them. Unselected LoRA modules remain silent and do not participate in the computation in the current computation step.

[0095] Step 32: Perform collaborative reasoning on the input data through the Transformer layer and the target low-rank adaptation module to generate output data.

[0096] The activated target low-rank adaptation module works in parallel or sequentially with the original Transformer layer, typically superimposing the LoRA output onto specific weights of the Transformer layer. The Transformer layer utilizes pre-trained general language knowledge, while the target low-rank adaptation module injects specialized corrections tailored to the specific optical network scenario. Both perform collaborative inference on the input data, fusing general logic with specialized experience, ultimately generating the output data of the current inference module.

[0097] The above process is repeated in each layer of the inference model. The data flow follows the path below: The first inference module receives the aligned input representation as its input data.

[0098] The next inference module receives the output data of the previous inference module as its input data, and then performs dynamic adaptation and selection through its own router.

[0099] After being processed by all the inference modules, the last inference module outputs the final high-level semantic features, which are then decoded to generate a text response, namely natural language text containing predictions, diagnoses, and suggestions.

[0100] This embodiment introduces a router before each Transformer layer, enabling the model to dynamically "select" the most suitable expert module from multiple candidate LoRA modules for activation based on the characteristics of the current input optical network data (such as different fault modes). This mechanism allows a single inference model to flexibly adapt to complex and varied fault scenarios in optical networks (handling both sudden outages and chronic aging), solving the problems of poor generalization ability and susceptibility to catastrophic forgetting in traditional static fine-tuning models, and significantly improving the accuracy of fault diagnosis and the robustness of the model.

[0101] Based on the inference model architecture of the above embodiment, the inference model is constructed through a series of specific fine-tuning steps to endow it with professional analytical capabilities for the optical network field.

[0102] Specifically, the model training server first performs the task of constructing training samples. These training samples mainly come from two parts of high-value data: One is the historical fault work order of optical network. This part of the data is extracted from the historical records of the actual optical network operation and maintenance system. It fully covers the text description of the entire life cycle from the triggering of fault phenomenon, the recording of the troubleshooting process, the location of the root cause to the implementation of the final repair solution, and provides a real operation and maintenance scenario mapping.

[0103] Secondly, there is the simulated expert dialogue. This part of the data is synthetic data generated by the more powerful Large Language Model (LLM) based on the preset optical network fault scenarios. It simulates the technical question-and-answer process between senior operation and maintenance experts and front-line personnel. In this way, it specifically supplements the complex fault reasoning chain that is relatively scarce in real data, thereby enriching the diversity and logical depth of the training samples.

[0104] After preparing the training samples containing rich semantic information, the server employs a parameter-efficient fine-tuning strategy to generate the final inference model.

[0105] During the training phase, the server strictly keeps all weight parameters of the pre-trained Transformer layer in the inference model frozen, meaning that these parameters are not updated during backpropagation. This fully preserves the original general language understanding and logical reasoning capabilities of the base model and significantly reduces the computational resources required for training.

[0106] Meanwhile, the server uses the training samples constructed above to calculate the loss function between the model output and the standard reference text, and uses the gradient descent algorithm to iteratively update and optimize the parameters of the newly added candidate low-rank adaptation module and the router in the model.

[0107] Through this training method, the router gradually learns to dynamically select the most suitable adapter module based on the characteristics of the input data, while the low-rank adapter module learns specific mapping rules for converting general language representations into optical network-specific diagnostic suggestions. Ultimately, the fully trained and converged candidate low-rank adapter modules and routers, together with the original frozen Transformer layer, form an inference model possessing specialized intelligence for the optical network domain.

[0108] As another possible implementation of the present invention Figure 4 This is a schematic diagram illustrating the performance data acquisition process during the preparation of the multimodal dataset provided by this invention. Figure 5 This is a flowchart illustrating the fine-tuning dataset construction process in the preparation of a multimodal dataset provided by this invention. (Combined with...) Figure 4 and Figure 5 As shown in the figure, this embodiment further elaborates on the process of collecting and constructing multimodal data used to train and support the above-mentioned inference model.

[0109] like Figure 4 As shown, the optical transport network (OTN) monitored by the optical network intelligent operation and maintenance server presents a layered architecture in physical topology, specifically encompassing the access network that directly connects to end users, the aggregation network responsible for traffic aggregation and distribution, and the backbone network that undertakes core transmission tasks. In the actual operation of the optical network, as... Figure 4 As indicated by the "Failure Occurred" label, network links or nodes inevitably experience failures or performance degradation. Servers collect multi-dimensional optical network performance data in real time through sensors and network management interfaces deployed at various network layers. This performance data forms the quantitative basis for fault prediction, specifically including but not limited to continuous time-series data such as the input / output optical power of each optical amplifier and transponder, board operating temperature, and laser bias current.

[0110] Building upon this, in order to construct a high-quality dataset that can endow the inference model with professional diagnostic intelligence, the server executes, for example... Figure 5 The data preparation process is shown below.

[0111] First, the server activates the data acquisition and aggregation module. This module acts as the data flow aggregation center, receiving the aforementioned real-time performance data from the optical network operating system and extracting unstructured text data from the operations and maintenance database. This text data mainly includes: fault work orders recording the entire process from fault occurrence to repair, event reports detailing network anomalies, and an operations and maintenance knowledge base storing equipment manuals and standard operating procedures.

[0112] To compensate for the scarcity of real-world fault samples in causal reasoning logic, the server further introduces a data augmentation mechanism based on the Large General Language Model (LLM). For example... Figure 5 As shown below, the server uses a general-purpose LLM as its knowledge engine and simulates the generation of a large amount of expert question-and-answer data through a pre-defined prompting process. This question-and-answer data, in the form of dialogue, simulates the process by which senior network experts perform logical deduction and decision-making for specific fault scenarios (such as abnormal attenuation of optical power).

[0113] Finally, the server cleans, aligns, and formats the fault tickets, event reports, and operation and maintenance knowledge bases derived from real-world operation and maintenance scenarios, along with the synthesized expert Q&A data, and integrates them to construct a high-quality fine-tuning dataset. This dataset not only contains the underlying physical parameter features of the optical network but also encompasses the top-level expert fault diagnosis logic. The server inputs this dataset into the subsequent LLM fine-tuning stage to perform domain-adaptive training on the pre-trained model, thereby ensuring that the final generated model possesses the necessary capabilities for... Figure 4 The ability to perceive the complex network topology shown has been mastered. Figure 5 The analysis logic for multi-source heterogeneous data is shown.

[0114] As another possible implementation of the present invention, in order to verify the effectiveness and superiority of the aforementioned optical network data processing method based on a large language model, this embodiment combines... Figures 6 to 10 The experimental data and practical application scenarios shown are explained in detail. Figure 6 This is a schematic diagram comparing the improved model provided by this invention with various existing open-source models in terms of training loss convergence characteristics. Figure 7 This is a schematic diagram comparing the performance indicators of the improved model provided by this invention with those of various existing open-source models in terms of multi-parameter trend prediction and fault diagnosis. Figure 8 This is a radar chart comparing the improved model provided by this invention with various existing open-source models in terms of multi-dimensional indicators of text generation quality and computational efficiency. Figure 9 This is a schematic diagram illustrating the performance analysis of the improved model provided by this invention in few-shot learning generalization under different proportions of training data. Figure 10 This is a schematic diagram of an interaction example of the optical network active defense framework based on a large language model provided by this invention in a real-world dialogue scenario.

[0115] The experimental environment in this embodiment is based on a real optical network operation and maintenance scenario, and the experimental data is collected from an optical transport network (OTN) operated by a certain operator. The dataset covers multi-dimensional, multi-parameter time-series performance data such as optical power, optical signal-to-noise ratio (OSNR), and bit error rate (BER), and is associated with annotated data containing semantic information, such as historical fault work orders and event reports. The test period fully covers the entire process of the network from normal operation, gradual performance degradation, to the final failure, ensuring the objectivity and comprehensiveness of the verification results.

[0116] During the model training phase, Figure 6 This paper presents a comparison of the convergence characteristics of the improved Llama 3.1-8B model proposed in this invention with several well-known open-source models (including Deepseek-R1-8B, Qwen3-8B, and Llama 3.2-3B) in terms of training loss. Figure 6 As shown, with the increase of training steps, the loss curve of the improved Llama 3.1-8B model decreases most rapidly and steadily, achieving the fastest convergence speed. This significant advantage is mainly attributed to the Time Series Label Embedding Module (TSEmb) based on Multilayer Perceptron (MLP) introduced into the model architecture. This module can efficiently extract the overall features of multivariate sequences, accelerating the model's learning process of the distribution patterns of optical network data.

[0117] During the model performance evaluation phase, the experiment conducted quantitative analysis from two dimensions: numerical trend prediction and fault state judgment. Figure 7 This study visually demonstrates the performance of different models in terms of accuracy, F1 score, and mean squared error (MSE). MSE is used to evaluate the data generation quality during the trend prediction phase, while ACC and F1 score are used to evaluate the classification accuracy during the fault identification phase. The results show that the improved Llama3.1-8B model achieves the best results across all metrics, with an accuracy of 0.9761, an F1 score of 0.9659, and the lowest mean squared error of 0.2012. This fully validates the model's superiority in handling multi-parameter collaborative prediction tasks, accurately capturing the complex relationships between parameters.

[0118] Furthermore, in order to evaluate the quality and efficiency of the natural language text generated by the model, Figure 8 A multi-dimensional comparison of indicators was conducted using radar charts. Analysis shows that the improved Llama3.1-8B model ranks first in both BLEU (71.15) and ROUGE-L (77.81) quality metrics, indicating that its generated diagnostic reports are semantically coherent and logically clear. Furthermore, in terms of runtime, this model has the shortest processing time among all comparable models, successfully achieving a synergistic optimization of generation accuracy and computational efficiency, meeting the stringent timeliness requirements of optical network operation and maintenance. Figure 9 This further demonstrates the model's outstanding performance in terms of data efficiency. Through few-shot learning generalization analysis, it is shown that using only 10% of the training data, the model maintains 97.14% of the baseline performance; when the data volume increases to 50%, the performance retention rate reaches 98.65%. This result indicates that the method of this invention has extremely strong generalization ability and can effectively address the challenge of scarce high-quality labeled data in actual operation and maintenance.

[0119] at last, Figure 10 This demonstrates specific application examples and effects of the proactive defense framework in real-world dialogue scenarios. Thanks to carefully designed prompt engineering and output constraints, the model can generate responses strictly according to the three-stage structure of "prediction-diagnosis-action." For example... Figure 10 As shown in the dialogue, when inputting historical data containing abnormal trends (such as optical power and bit error rate data) and corresponding prompts, the model first outputs accurate predictions of key parameter values ​​for the next 7 days, identifying abnormal patterns such as "both input and output optical power will gradually decrease" and "the bit error rate is expected to exceed the threshold." Next, the model performs root cause analysis based on domain knowledge, inferring possible causes such as "laser aging, connector contamination, or gain loss." Finally, the model outputs specific and actionable proactive defense recommendations, such as "recommend performing preventative checks on the transmitter and optical connectors within 24 hours," "perform fiber cleaning and re-insertion," and "set automatic alarm triggers." Practical verification shows that these generated defense recommendations can be directly converted into maintenance work orders or automated execution scripts, achieving a complete closed loop from intelligent prediction to actual operation. Compared to traditional methods, this solution can identify potential faults 7 time units in advance, effectively suppressing the "illusion" problem of large models and ensuring that each recommendation has clear data support and logical basis, thus achieving a fundamental shift in optical network fault management from passive response to proactive defense.

[0120] Reference Figure 11 , Figure 11 This is a schematic diagram of the structure of the optical network data processing system based on a large language model provided by the present invention. The system includes: The first processing module is used to map each historical numerical sequence in the time series data of the optical network to a feature embedding vector, and generate a labeled embedding sequence based on all feature embedding vectors; The second processing module is used to construct text prompt words containing task prompt information, and modally align the token embedding sequence with the text features of the text prompt words to generate an aligned input representation; The third processing module is used to input the aligned input representation into the pre-trained inference model and obtain the text response output by the inference model that is associated with the time series data; wherein, the inference model is obtained by fine-tuning the large language model.

[0121] It should be noted that the optical network data processing system based on a large language model provided by the present invention can execute the optical network data processing method based on a large language model of any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0122] Figure 12 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 12 As shown, the electronic device may include a processor 1210, a communications interface 1220, a memory 1230, and a communication bus 1240. The processor 1210, communications interface 1220, and memory 1230 communicate with each other via the communication bus 1240. The processor 1210 can call logical instructions in the memory 1230 to execute the optical network data processing method based on a large language model provided in the above embodiments.

[0123] Furthermore, the logical instructions in the aforementioned memory 1230 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0124] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer is able to execute the optical network data processing method based on the large language model provided in the above embodiments.

[0125] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the optical network data processing method based on the large language model provided in the above embodiments.

[0126] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0127] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.< / t> < / n>

Claims

1. A method for optical network data processing based on a large language model, characterized in that, include: Each historical numerical sequence in the time-series data of the optical network is mapped to a feature embedding vector, and a labeled embedding sequence is generated based on all the said feature embedding vectors; Construct text prompt words containing task prompt information, and modally align the embedded sequence of the markers with the text features of the text prompt words to generate an aligned input representation; The aligned input representation is fed into a pre-trained inference model to obtain the text response output by the inference model that is associated with the time series data; wherein the inference model is obtained by fine-tuning a large language model.

2. The optical network data processing method based on a large language model according to claim 1, characterized in that, The step of modally aligning the embedded sequence of the markers with the text features of the text prompt words to generate an aligned input representation includes: The cross-attention mechanism is used to process the labeled embedding sequence and the text features to generate an aligned input representation; Specifically, for each attention head of the cross-attention mechanism, a query vector is generated based on the labeled embedding sequence, and a key vector and a value vector are generated based on the text features; The output of the current attention head is generated based on the query vector, the key vector, and the value vector. The outputs of all the attention heads are concatenated and then linearly projected to generate the aligned input representation.

3. The optical network data processing method based on a large language model according to claim 2, characterized in that, The aligned input representation is calculated using the following formula: ； ； In the formula, K represents the number of attention heads. For the output of the k-th attention head, Let k be the query vector of the attention head. Let k be the key vector of the attention head. Let T be the value vector of the k-th attention head, and let T be the transpose sign. Let be the dimension of the feature vector in the attention head. For normalized exponential functions, This represents the aligned input. For concatenation functions, This is for outputting the projection matrix.

4. The optical network data processing method based on a large language model according to claim 1, characterized in that, The process of mapping each historical numerical sequence in the time-series data of the optical network to a feature embedding vector includes: For each network variable in the time series data, obtain the complete historical value sequence within a preset time window; The historical numerical sequence is processed by a fully connected projection using a multilayer perceptron. The historical numerical sequence is compressed and mapped into a fixed-dimensional feature embedding vector, wherein the feature embedding vector represents the overall dynamic trend of the network variables within the preset time window.

5. The optical network data processing method based on a large language model according to claim 1, characterized in that, The inference model includes multiple inference modules and multiple candidate low-rank adaptation modules; each inference module includes a Transformer layer and a router set before the Transformer layer; The step of feeding the aligned input representation into a pre-trained inference model to obtain the text response output by the inference model associated with the time series data includes: For any one of the inference modules, the router calculates the matching score between the input data and each of the candidate low-rank adaptation modules, and based on the matching score, selects the matching target low-rank adaptation module from the plurality of candidate low-rank adaptation modules and activates the target low-rank adaptation module. The input data is collaboratively inferred by the Transformer layer and the target low-rank adaptation module to generate output data. The input data of the first inference module is the aligned input representation; the input data of the subsequent inference module is the output data of the previous inference module; and the output data of the last inference module is the text response.

6. The optical network data processing method based on a large language model according to claim 5, characterized in that, The inference model was obtained through the following fine-tuning steps: Acquire training samples, which include historical fault tickets of optical networks and simulated expert dialogues; the simulated expert dialogues are generated by a general large language model. Keeping the parameters of the Transformer layer frozen, the parameters of the candidate low-rank adaptation module and the router are updated and trained using the training samples to generate the inference model.

7. A data processing system for optical networks based on a large language model, characterized in that, include: The first processing module is used to map each historical numerical sequence in the time series data of the optical network to a feature embedding vector, and generate a labeled embedding sequence based on all the said feature embedding vectors; The second processing module is used to construct text prompt words containing task prompt information, and modally align the embedded sequence of the markers with the text features of the text prompt words to generate an aligned input representation; The third processing module is used to input the aligned input representation into a pre-trained inference model to obtain the text response output by the inference model that is associated with the time series data; wherein the inference model is obtained by fine-tuning a large language model.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the optical network data processing method based on a large language model as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the optical network data processing method based on a large language model as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the optical network data processing method based on a large language model as described in any one of claims 1 to 6.

Images