A key information extraction method and system based on multimodal model

Abstract

The present application provides a key information extraction method and system based on a multimodal model. Among them, the solution integrates building structure stress and optical cable topology data, and predicts the probability of node failure by correlating structural deformation with optical signal attenuation characteristics through a multimodal model. Screen high failure risk nodes and plan isolated backup optical cable paths, use spatial diversity technology to synchronize the main and backup signals, and generate a redundant signal set by cross-modal alignment. Divide the redundant data into slices according to the geographic grid, and generate a multimodal hash identifier based on the wavelength of the optical signal of the remote node and the regional pressure resistance parameters. When a node fails, match the data slices in the high pressure resistance area according to the hash identifier, and reconstruct key information using the multi-path time-frequency complementary features to achieve network fault self-healing and intelligent redundant recovery. The technical solution provided by the present application realizes the intelligence and efficiency of communication network failure prediction and redundant recovery, and significantly improves network stability and business continuity.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a key information extraction method and system based on a multimodal model. Background Art

[0002] In modern buildings and communication infrastructure, the dynamic deformation of building structures and the stability of optical cable networks are crucial to the transmission and storage of critical information. Especially in high-rise buildings, industrial facilities or earthquake-prone areas, the stress distribution and deformation of building structures will directly affect the communication quality of optical cable networks, and even cause node failure or data loss.

[0003] The extraction of key information in the prior art usually relies on the multi-path transmission and redundant storage mechanism of optical signals. Specifically, key information is transmitted through multiple paths of the optical cable network, and redundant copies are stored in multiple nodes. When a path or node failure is detected, the system extracts redundant information from other paths or nodes to restore key data.

[0004] However, the existing technology does not fully consider the impact of dynamic deformation of building structures on optical cable networks, resulting in reduced efficiency in extracting and recovering key information when the building deforms or external pressure changes. Summary of the invention

[0005] The embodiments of the present application provide a key information extraction method and system based on a multimodal model to solve the problem of low efficiency in extraction and recovery of key information in the prior art.

[0006] In a first aspect, an embodiment of the present application provides a key information extraction method based on a multimodal model, comprising:

[0007] Acquire building structure stress distribution data and optical cable layout topology data, wherein the building structure stress distribution data includes dynamic deformation parameters of each area of ​​the building, and the optical cable layout topology data includes the spatiotemporal correlation between the physical position of the communication node and the optical signal transmission path;

[0008] Generate a failure probability of a communication node by fusing the dynamic deformation parameter with the historical attenuation characteristics of the optical signal transmission path in the spatiotemporal correlation relationship through a multimodal model;

[0009] Based on the failure probability, a backup optical cable path physically isolated from the node with high failure probability is extracted from the time-space correlation relationship, the optical signals of the primary and backup paths are synchronously received through the space diversity receiving technology, and the time-frequency domain features of the primary and backup paths are cross-modally aligned in combination with the multimodal model to generate a redundant signal set;

[0010] Divide the key information in the redundant signal set into data fragments according to the node geographical distribution rule, associate the data fragments with the optical signal wavelength characteristics of the remote nodes and the building structure compressive strength parameters through the multimodal model, and generate a multimodal hash identifier;

[0011] When failure of the target communication node is detected, data fragments having the same optical signal wavelength characteristics and located in a high compressive strength area are matched from the remote node according to the multimodal hash identifier, and the key information is reconstructed based on the complementary characteristics of the multipath signals in the redundant signal set.

[0012] Optionally, matching data fragments having the same optical signal wavelength characteristics and located in a high compressive strength area from the remote node according to the multimodal hash identifier, and reconstructing the key information based on the complementary characteristics of multipath signals in the redundant signal set, including:

[0013] Establishing a distributed hash mapping table in a remote node according to the multimodal hash identifier, performing a neighbor search on the hash mapping table, and screening and matching data fragments having the same optical signal wavelength characteristics based on the neighbor search;

[0014] During the data shard matching process, the pressure sensor data of the remote nodes is collected in real time, and a dynamic matrix of node compressive strength is constructed. The compressive coefficient of each node is calculated through the dynamic matrix of node compressive strength, and nodes with large compressive coefficients are retained. Based on the nodes with large compressive coefficients, data shards located in the high compressive strength area are screened out;

[0015] A space-frequency joint dictionary is established according to the complementary characteristics of the multipath signals in the redundant signal set, signal reconstruction error minimization and energy convergence conditions are output based on the space-frequency joint dictionary, and key information is output after convergence verification based on the signal reconstruction error minimization and energy convergence conditions.

[0016] Optionally, it is characterized in that the step of establishing a space-frequency joint dictionary according to the complementary characteristics of the multipath signals in the redundant signal set, outputting a signal reconstruction error minimization and energy convergence condition based on the space-frequency joint dictionary, and outputting key information after convergence verification based on the signal reconstruction error minimization and the energy convergence condition, comprises:

[0017] Extract the spatiotemporal propagation characteristics of the complementary features of multipath signals from the redundant signal set, eliminate the phase ambiguity of the sensor array through space-frequency joint projection, and output the standardized multipath complementary feature tensor;

[0018] The multi-path complementary feature tensor is optimized through orthogonal constraints to optimize the incoherence between basis functions and generate a space-frequency joint dictionary with path difference compensation characteristics.

[0019] The reconstruction error objective function is constructed based on the space-frequency joint dictionary, and the second-order derivative matrix of the error function is optimized through the gradient descent strategy to output the signal reconstruction error minimization and energy convergence conditions;

[0020] Based on the signal reconstruction error minimization and energy convergence conditions, a stable convergence path is screened through a double verification mechanism to output key information to resist multipath interference.

[0021] Optionally, it is characterized in that the failure probability of the communication node is generated by fusing the dynamic deformation parameter with the historical attenuation characteristics of the optical signal transmission path in the spatiotemporal correlation relationship through a multimodal model, including:

[0022] The dynamic deformation parameter sequence and the spatiotemporal correlation are fused through a multimodal model, and the weight proportions of the two types of features in the optical signal transmission path are dynamically allocated to generate a fusion feature matrix.

[0023] Performing multi-scale feature decomposition on the fused feature matrix to extract a multi-scale feature matrix containing steady-state attenuation components and transient jump components of different time granularities;

[0024] Based on the multi-scale feature matrix, the feature contribution of each time granularity is adjusted by a dynamic weight optimization algorithm to generate an optimized fusion feature matrix;

[0025] Performing state transition analysis on the optimized fusion feature matrix to calculate a piecewise linear function of the historical state attenuation feature of the optical signal transmission path;

[0026] Based on the piecewise linear function, the numerical interval of the optimized fusion feature matrix is ​​mapped to the initial failure probability, the initial failure probability distribution is subjected to physical constraint verification, and the failure probability of the communication node is output.

[0027] Optionally, it is characterized in that a backup optical cable path physically isolated from a high failure probability node is extracted from the time-space association relationship, the primary and backup path optical signals are synchronously received through a space diversity reception technology, and the time-frequency domain features of the primary and backup path optical signals are cross-modally aligned in combination with the multimodal model to generate a redundant signal set, including:

[0028] Based on the spatiotemporal association, all candidate optical cable paths are traversed to select candidate optical cable paths that meet the minimum geographical isolation distance with the high failure probability node as backup optical cable paths, and candidate optical cable paths that have overlapping routes with the main optical cable path are excluded;

[0029] Diversity receiving devices are deployed at the receiving ends of the main optical cable path and the backup optical cable path to achieve synchronous reception of optical signals of the main and backup paths through spatial diversity receiving technology;

[0030] The time domain waveform features and frequency domain spectrum features are extracted from the primary and backup path optical signals after noise suppression respectively and input into the multimodal model. The cross-modal alignment method is used to fuse the time and frequency domain features of the primary and backup path signals to generate a redundant signal set.

[0031] Optionally, it is characterized in that the data fragment is associated with the optical signal wavelength characteristics of the remote node and the compressive strength parameters of the building structure through the multimodal model to generate a multimodal hash identifier, including:

[0032] Based on the multimodal model, the optical signal of the remote node is subjected to frame and window processing, wherein the frame and window processing process is adaptively adjusted according to the building vibration frequency to construct an optical signal feature sequence;

[0033] Mapping the compressive strength parameters of the building structure to a three-dimensional space grid coordinate system, calculating the dynamic compressive strength attenuation factor of each grid unit based on the three-dimensional space grid coordinate system, and generating a compressive strength characteristic tensor based on the compressive strength attenuation factor strength;

[0034] A mapping relationship table between the optical signal feature sequence and the compressive strength feature tensor is established, and a joint feature matrix including wavelength and compressive strength coupling weights is generated based on the mapping relationship table;

[0035] The correlation between the joint features is eliminated to generate a multimodal hash identifier with anti-collision characteristics.

[0036] Optionally, it is characterized in that a state transition analysis is performed on the optimized fusion feature matrix to calculate a piecewise linear function of the historical state attenuation feature of the optical signal transmission path, including:

[0037] Based on the position parameters in the fusion feature matrix, state features in the local neighborhood are extracted through a preset neighborhood set to generate a neighborhood state feature set;

[0038] Performing time-varying weight assignment on the neighborhood state feature set, introducing an attenuation coefficient related to the time step for each neighborhood state, and dynamically adjusting the nonlinear attenuation gradient of the attenuation coefficient through a piecewise linear interpolation algorithm;

[0039] State transition analysis is performed on the neighborhood state feature set after the attenuation coefficient is adjusted, the historical state is divided into multiple intervals, and the boundary value jump parameters across the intervals are calculated to generate a piecewise linear function of the attenuation characteristics of the historical state of the optical signal transmission path.

[0040] Optionally, it is characterized in that, based on the piecewise linear function, the numerical interval of the optimized fusion feature matrix is ​​mapped to the initial failure probability, the initial failure probability distribution is physically constrained and verified, and the failure probability of the communication node is output, including:

[0041] Dividing the numerical interval of the fusion feature matrix into multiple sub-intervals based on the piecewise linear function, performing linear transformation on the numerical values ​​in the sub-intervals, and generating an initial failure probability;

[0042] The initial failure probability is dynamically adjusted according to the deviation between the actual observation value and the initial probability, the rationality of the distribution is verified, and the physical constraint verification of the initial failure probability distribution is performed;

[0043] The failure probability after physical constraint verification is associated with the dynamic parameters of the communication node, and based on the failure probability data between different communication nodes, a communication node failure probability with a unified benchmark is generated.

[0044] Optionally, it is characterized in that all candidate optical cable paths are traversed based on the spatiotemporal association relationship, candidate optical cable paths that meet the minimum geographical isolation distance with the high failure probability node are screened out as backup optical cable paths, and candidate optical cable paths that have overlapping routes with the main optical cable path are excluded, including:

[0045] Based on the time-space correlation relationship, each candidate optical cable path in the candidate optical cable path set is subjected to transmission delay equalization processing, and the delay difference between the paths is eliminated by dynamically adjusting the optical fiber dispersion compensation parameters, so as to generate a set of candidate optical cable path parameters after delay equalization;

[0046] Performing geographic coordinate mapping on the candidate optical cable path parameter set after delay equalization, calculating the geographic isolation distance between adjacent optical cable paths, and screening out backup optical cable paths that meet the minimum geographic isolation distance with high failure probability nodes;

[0047] The optical cable links of the main path are extracted. If there are optical cable links identical to the main path in the backup optical cable path after noise suppression, it is determined to be a route overlapping path, and the candidate optical cable paths of the overlapping routes are excluded.

[0048] In a second aspect, an embodiment of the present application provides a key information extraction system based on a multimodal model, comprising:

[0049] An acquisition module is used to acquire building structure stress distribution data and optical cable layout topology data, wherein the building structure stress distribution data includes dynamic deformation parameters of each area of ​​the building, and the optical cable layout topology data includes the spatiotemporal correlation between the physical position of the communication node and the optical signal transmission path;

[0050] A generation module, which generates a failure probability of a communication node by fusing the dynamic deformation parameter with the historical attenuation characteristics of the optical signal transmission path in the spatiotemporal correlation relationship through a multimodal model;

[0051] A generation module, based on the failure probability, extracts a backup optical cable path physically isolated from the high failure probability node from the time-space association relationship, synchronously receives the primary and backup path optical signals through the space diversity reception technology, and performs cross-modal alignment on the time-frequency domain features of the primary and backup path optical signals in combination with the multimodal model to generate a redundant signal set;

[0052] A generation module is provided to divide the key information in the redundant signal set into data fragments according to the node geographical distribution rule, and to associate the data fragments with the optical signal wavelength characteristics of the remote nodes and the compressive strength parameters of the building structure through the multimodal model to generate a multimodal hash identifier;

[0053] The reconstruction module, when detecting the failure of the target communication node, matches the data fragments having the same optical signal wavelength characteristics and located in the high compressive strength area from the remote node according to the multimodal hash identifier, and reconstructs the key information based on the complementary characteristics of the multipath signals in the redundant signal set.

[0054] In an embodiment of the present application, building structure stress distribution data and optical cable layout topology data are obtained, wherein the building structure stress distribution data includes dynamic deformation parameters of each area of ​​the building, and the optical cable layout topology data includes the spatiotemporal correlation between the physical position of the communication node and the optical signal transmission path; the dynamic deformation parameters are fused with the historical attenuation characteristics of the optical signal transmission path in the spatiotemporal correlation through a multimodal model to generate a failure probability of the communication node; based on the failure probability, a backup optical cable path physically isolated from a node with a high failure probability is extracted from the spatiotemporal correlation, and the optical signals of the main and backup paths are synchronously received through a spatial diversity reception technology, and the failure probability of the communication node is generated. The multimodal model is used to perform cross-modal alignment on the time-frequency domain characteristics of the primary and backup path optical signals to generate a redundant signal set; the key information in the redundant signal set is divided into data slices according to the node geographical distribution rules, and the data slices are associated with the optical signal wavelength characteristics of the remote nodes and the compressive strength parameters of the building structure through the multimodal model to generate a multimodal hash identifier; when the failure of the target communication node is detected, the data slices with the same optical signal wavelength characteristics and located in the high compressive strength area are matched from the remote nodes according to the multimodal hash identifier, and the key information is reconstructed based on the complementary characteristics of the multipath signals in the redundant signal set.

[0055] The technical solution of this application has the following beneficial effects:

[0056] The dynamic deformation parameter sequence and the spatiotemporal correlation are fused through the time dimension alignment module, and the weight proportions of the two types of features in the optical signal transmission path are dynamically allocated to generate a fused feature matrix. State transition analysis is performed on the fused feature matrix, and the piecewise linear function of the historical state attenuation characteristics of the optical signal transmission path is calculated. Based on the piecewise linear function, the numerical interval of the fused feature matrix is ​​mapped to the initial failure probability, the initial failure probability distribution is modified and verified, and the failure probability of the communication node is output.

[0057] Furthermore, through time dimension alignment and dynamic weight allocation, the deep fusion of dynamic deformation parameters and spatiotemporal correlation is achieved to generate a high-precision fusion feature matrix; based on state transition analysis and piecewise linear function, the historical attenuation characteristics of the optical signal transmission path are accurately calculated, and through probability mapping and verification, the failure probability of the communication node is output, providing a reliable basis for subsequent path selection and redundant recovery.

[0058] These and other aspects of the present application will become more clearly understood in the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0060] Figure 1 A flowchart of a key information extraction method based on a multimodal model provided by the present application is shown;

[0061] Figure 2 A structural schematic diagram of a key information extraction system based on a multimodal model provided by the present application is shown. DETAILED DESCRIPTION

[0062] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0063] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0064] This application constructs an integrated solution for communication network failure prediction and redundancy recovery through multimodal data fusion and intelligent analysis. First, the building structure stress distribution data and optical cable layout topology data are integrated, and the dynamic deformation parameters are integrated with the historical attenuation characteristics of the optical signal transmission path using a multimodal model to generate the failure probability of the communication node; secondly, the physically isolated backup optical cable paths are screened based on the failure probability, and the main and backup path optical signals are synchronized using spatial diversity reception technology, and a redundant signal set is generated through cross-modal alignment; then, the key information of the redundant signal is divided into data fragments according to the geographical distribution rules, and a multimodal hash identifier is generated by combining the wavelength characteristics of the optical signal and the compressive strength parameters of the building structure; finally, when the target node fails, the multimodal hash identifier is used to match the data fragments in the high compressive strength area, and the key information is reconstructed through the complementary characteristics of the multipath signal to achieve intelligent redundancy recovery and efficient fault tolerance of the communication network.

[0065] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0066] Figure 1 A flowchart of a key information extraction method based on a multimodal model is provided for an embodiment of the present application, such as Figure 1 As shown, the method includes:

[0067] 101. Acquire building structure stress distribution data and optical cable layout topology data, wherein the building structure stress distribution data includes dynamic deformation parameters of each area of ​​the building, and the optical cable layout topology data includes a spatiotemporal correlation between a physical position of a communication node and an optical signal transmission path;

[0068] In this step, the stress distribution data of the building structure are dynamic deformation parameters collected through a distributed optical fiber sensor network, including strain gradient (με / mm), vibration energy spectrum density (J / Hz), load distribution (kN / m²), etc., reflecting the real-time mechanical state of each area of ​​the building.

[0069] The optical cable layout topology data describes the physical location (three-dimensional coordinates) of the communication node and the spatial and temporal relationship of the optical signal transmission path, including parameters such as the number of path hops, transmission delay (μs), and optical signal-to-noise ratio (OSNR).

[0070] In the embodiment of the present application, a fiber grating sensor array is deployed in the load-bearing structure of the building (such as beam-column joints and shear walls), and dynamic deformation data is collected at a sampling frequency of 200Hz, and the strain gradient distribution is calculated through a finite element analysis model. Use an optical time domain reflectometer (OTDR) to scan the optical cable path, record the link loss (dB / km) and delay jitter (μs²) between nodes, and combine the Beidou positioning system to obtain the precise coordinates of the nodes (error ≤ 5cm). After the two types of data are aligned by GPS timestamps, a spatiotemporal correlation database is constructed, and each data point contains position, time, strain gradient and optical signal parameters.

[0071] A certain financial center building deployed a fiber optic sensor network in the core area, and monitored the beam body strain gradient of 0.1με / mm and the main vibration frequency of 12Hz in real time. At the same time, the OTDR determined that the optical cable path loss between node A (coordinates X1, Y1, Z1) and node B (coordinates X2, Y2, Z2) was 0.6dB / km and the delay was 3.1ms. After the data was aligned with the timestamp, a joint database of building stress and optical cable topology was formed.

[0072] 102. Generate a failure probability of a communication node by fusing the dynamic deformation parameter with the historical attenuation characteristics of the optical signal transmission path in the spatiotemporal correlation relationship through a multimodal model;

[0073] In this step, the multimodal model is a dual-channel deep learning framework based on the graph convolutional network (GCN), which inputs mechanical modes (strain gradient, vibration energy) and communication modes (optical signal attenuation rate, delay jitter), and outputs node failure probability (value in the range of 0-1).

[0074] The historical attenuation characteristics include dynamic parameters such as the accumulated optical signal path loss (unit: dB / month) and the dispersion coefficient change trend (ps / nm / km·day).

[0075] In the embodiment of the present application, the strain gradient change rate (Δε / Δt, reflecting deformation acceleration) and vibration energy spectrum entropy (quantifying the randomness of vibration energy distribution) are extracted, the path loss slope (dB / day, reflecting the signal attenuation rate) and delay fluctuation variance (μs², characterizing transmission stability) are calculated, and a dual-branch GCN model is constructed. The mechanical branch processes node strain data, and the communication branch processes optical signal parameters. Cross-attention mechanism is used to achieve cross-modal feature interaction. The final output layer uses Sigmoid function to map failure probability, and the loss function uses weighted cross entropy to give higher weights to nodes in high strain areas (≥0.15με / mm). The training data contains historical failure events (such as optical cable breakage, node downtime) and corresponding modal features.

[0076] The strain gradient change rate of node C (located in the deformation-sensitive area of ​​the 80th layer) reached 0.2με / mm·day in the past 15 days, and the corresponding optical signal path loss slope rose to 0.45dB / day, and the delay fluctuation variance was 12μs². The model found that the mechanical and communication characteristics were highly correlated (correlation coefficient 0.85) through the cross-attention mechanism, and predicted that the failure probability of node C was 0.88 (extremely high risk).

[0077] 103. Based on the failure probability, extract a backup optical cable path that is physically isolated from the node with high failure probability from the time-space correlation relationship, synchronously receive the primary and backup path optical signals through the space diversity receiving technology, and perform cross-modal alignment on the time-frequency domain features of the primary and backup path optical signals in combination with the multimodal model to generate a redundant signal set;

[0078] In this step, the physically isolated backup path is an optical cable route that has no physical overlap with high-failure nodes, such as flying cables outside the building or underground redundant pipelines, with a hop count of ≤3 and a delay difference of ≤2ms.

[0079] Spatial diversity reception technology uses a multiple-input multiple-output (MIMO) antenna array to receive primary and backup path signals and improves the signal-to-noise ratio through maximum ratio combining (MRC).

[0080] In the embodiment of the present application, for nodes with a failure probability ≥ 0.8, an improved A The algorithm searches for backup paths and avoids high-risk areas with strain gradients ≥ 0.1με / mm. The planning results must meet the hop count constraint (e.g., 3 hops for the main path → 2 hops for the backup path). The signals of the main and backup paths are received through an 8×8 MIMO array, and the minimum mean square error (MMSE) equalizer is used to eliminate multipath interference. The delay synchronization accuracy is controlled within ±0.5μs. The main and backup signals are subjected to short-time Fourier transform (STFT, window length 512ms, overlap rate 80%) to generate a time-frequency energy matrix. The time-frequency characteristics of the main and backup signals are aligned to eliminate the phase offset caused by path differences and generate a redundant signal set (signal-to-noise ratio improvement ≥ 6dB).

[0081] The main path of node C passes through the high-risk area on the 80th floor, and the backup path is redirected to the flying cable outside the building (2 hops, 1.8ms delay difference). After the main and backup signals are received by MIMO, STFT shows that the energy of the main path drops by 10dB in the 8Hz frequency band (resonance with the structural vibration frequency). The backup path is aligned by DTW to generate a redundant set, and the signal-to-noise ratio is improved from 15dB to 21dB after merging.

[0082] 104. Divide the key information in the redundant signal set into data fragments according to the node geographical distribution rule, associate the data fragments with the optical signal wavelength characteristics of the remote nodes and the building structure compressive strength parameters through the multimodal model, and generate a multimodal hash identifier;

[0083] In this step, data sharding is to divide the key information (such as routing tables, user session data) in redundant signals according to geographic grids (50m×50m×10 layers), and each shard is bound to compressive strength (unit: MPa) and optical wavelength characteristics (nm-level accuracy).

[0084] The multimodal hash identifier is a 64-bit identifier consisting of a wavelength hash (32 bits) and a compressive strength hash (32 bits), with an additional CRC-8 checksum (8 bits).

[0085] In the embodiment of the present application, redundant signals are divided according to the building space grid. For example, the slice G-20 corresponds to the 80th-90th floor and the east side area. The main wavelength (1550.12±0.01nm) and spectral width (0.02nm) are extracted from the slice optical signal, and the concrete compressive strength of the grid (such as C60 corresponds to 60MPa) is associated at the same time. The wavelength feature is input into the local sensitive hash (LSH), and the compressive strength parameter is input into the consistent hash (CH) to generate a 64-bit composite hash code. The Hamming distance threshold (≤3) and CRC check are used to exclude duplicate hashes to ensure the uniqueness of the slice.

[0086] Shard S-15 corresponds to grid G-20 (compressive strength 60MPa), whose main wavelength of optical signal is 1550.12nm and spectrum width is 0.02nm. The model generates hash mark "0x3A7D5F..." through LSH and CH, and stores it in the blockchain account book to support subsequent fast retrieval.

[0087] 105. When failure of the target communication node is detected, data fragments having the same optical signal wavelength characteristics and located in a high compressive strength area are matched from the remote node according to the multimodal hash identifier, and the key information is reconstructed based on the complementary characteristics of the multipath signals in the redundant signal set.

[0088] In this step, the complementary characteristics of the multipath signals are the complementarity of the primary and backup path signals in the time domain (phase difference) and frequency domain (energy distribution), such as compensation of the backup path high-frequency signal when the primary path low-frequency energy is lost.

[0089] In an embodiment of the present application, slices with the same wavelength (±0.1nm) and compressive strength ≥50MPa are screened according to the target node coordinates and hash identifiers. Weights are dynamically allocated based on the complementarity of the time-frequency domain (such as a main path weight of 0.6 and a backup path weight of 0.4), and the signals are aligned and superimposed using a phase synchronization algorithm. Reed-Solomon coding is used to correct transmission errors, and the packet loss rate is reduced from 1.2% to 0.05%. If multiple slice data conflict, the optimal solution is selected according to the hash identifier priority (the higher the compressive strength, the higher the priority).

[0090] After node C fails, the high-pressure area (grid G-20, 60MPa) is matched with slice S-15 (wavelength 1550.12nm). The main path signal is lost in the 8Hz frequency band (due to structural resonance), and the backup path signal is restored to a data packet integrity rate of 99.95% through phase compensation (compensation amount π / 4) and weighted fusion (weight 0.6:0.4).

[0091] In summary, this solution realizes an integrated closed loop of communication node failure prediction and intelligent redundant recovery by deeply integrating building mechanics and optical communication data. In steps 101-102, high-precision sensors and multimodal models increase the accuracy of node failure prediction to 92%; steps 103-104 ensure the high availability of data sharding in extreme scenarios (99.9% data integrity in areas with compressive strength ≥ 50MPa) through physical isolation paths and multimodal hash identification; and step 105's multipath complementary reconstruction technology compresses the recovery time to within 1.5 seconds. The implementation example shows that this solution can reduce network interruption time by 80% and redundant resource consumption by 40% in super-high-rise buildings, and is suitable for high-risk scenarios such as data centers in earthquake zones and urban landmark buildings.

[0092] In order to improve the accuracy and engineering interpretability of the prediction of the failure probability of communication nodes, a cross-modal attention mechanism is first designed to dynamically fuse the dynamic deformation parameters of the building structure (such as the strain gradient change rate) and the spatiotemporal correlation of the optical signal (such as the path loss slope). The time series of the two types of features are extracted through a sliding time window, and a gated weight allocation strategy is used to generate a fusion feature matrix. Then, the multi-scale wavelet packet decomposition technology is introduced to separate the fusion feature matrix into steady-state attenuation components (reflecting the long-term structural deformation trend) and transient jump components (capturing sudden optical signal degradation) of different time granularities to construct a multi-scale feature matrix. Based on the feature contribution evaluation model, a dynamic weight optimization algorithm (such as a reinforcement learning framework) is designed to adaptively adjust the weights of the features of each time granularity to generate an optimized fusion matrix under physical constraints. The hidden Markov model (HMM) is further combined to analyze the historical state transition law of the optical signal, and a piecewise linear attenuation function is fitted to describe the path degradation trajectory. Finally, the optimized features are mapped to the initial failure probability through a Bayesian probability framework, and the building structure mechanical parameters (such as the node stiffness threshold) are superimposed for physical verification to output the communication node failure probability that meets the actual engineering constraints. In some embodiments, the failure probability of a communication node is generated by fusing the dynamic deformation parameter with the historical attenuation characteristics of the optical signal transmission path in the spatiotemporal correlation relationship through a multimodal model, including:

[0093] 201. The dynamic deformation parameter sequence and the spatiotemporal correlation are fused through a multimodal model, and the weight proportions of the two types of features in the optical signal transmission path are dynamically allocated to generate a fusion feature matrix;

[0094] In step 201, the dynamic deformation parameter sequence is the time series data of the mechanical parameters of the building structure collected in real time by the high-density sensor network, such as the strain gradient (unit: microstrain / mm), vibration energy spectrum density (reflecting the structural resonance intensity), and load distribution (unit: kilonewton / square meter) of the key nodes of the building (such as beams, columns, and shear walls). The spatiotemporal correlation relationship is the physical topological parameter of the optical communication network, including the three-dimensional coordinates of the node, the length of the optical cable path, the number of hops (the number of relay nodes), the delay jitter (reflecting the stability of signal transmission), the optical signal-to-noise ratio (measurement of signal quality) and other spatiotemporal dynamic parameters. After the fusion feature matrix aligns the mechanical deformation features with the communication parameters according to the time window, the unified feature expression generated by dynamic weight allocation is used to describe the comprehensive state of the node in a specific time period.

[0095] In the embodiment of the present application, a cross-modal attention mechanism is used to achieve deep fusion of two types of features. The dynamic deformation parameters are collected at a high sampling rate (such as 200Hz) and divided into continuous sequences through sliding time windows (such as 60-second windows with a step size of 10 seconds). Each window contains data such as strain gradients and vibration energy at multiple time points. Optical signal parameters (such as delay jitter and path loss) are synchronously divided into the same time window, and timestamp alignment is ensured by a high-precision clock synchronization protocol (such as PTP), and the error is controlled at the microsecond level. The aligned data is organized by node ID to form an original data set containing four-dimensional information of time, space, mechanics, and communication. A dual-channel gating network is designed to process mechanical features (such as strain gradient sequences) and communication features (such as delay jitter sequences) respectively. By analyzing the strain gradient change trend (such as continuous rise or mutation), combining historical data to learn its impact on node stability, a weight of 40%-60% is dynamically assigned. According to parameters such as delay jitter variance and optical signal-to-noise ratio degradation rate, the communication link reliability score is calculated, and the weight is complementary to the mechanical features (60%-40%). The weight distribution is adjusted in real time. For example, when the strain gradient of a node suddenly increases, the mechanical weight is automatically increased to 55%, and the communication weight is correspondingly reduced to 45%. The weighted mechanical and communication features are spliced ​​according to the time window to generate a three-dimensional fusion feature matrix with the dimensions of the number of nodes × the number of time windows × the number of features (such as 100 nodes × 100 windows × 2 types of features). Each matrix element represents the comprehensive state of the node in a specific time window.

[0096] 202. Perform multi-scale feature decomposition on the fused feature matrix to extract a multi-scale feature matrix containing steady-state attenuation components and transient jump components of different time granularities;

[0097] In step 202, the steady-state attenuation component reflects the long-term slow change trend of optical signal path loss and structural deformation, and the transient jump component captures short-term drastic fluctuations caused by sudden events (such as earthquake shock, equipment overload), such as a sudden increase in delay in seconds or an abnormal strain in minutes.

[0098] In the embodiment of the present application, wavelet packet decomposition technology is used to extract multi-scale information from the fusion features, and the Daubechies wavelet basis function (such as DB4) is selected because of its balance in time-frequency localization characteristics, which is suitable for capturing transient and steady-state components in engineering signals. Perform 3-layer wavelet packet decomposition to decompose the fusion features of each time window into 8 sub-bands (scales), covering the frequency range of 0-4Hz. Select the lowest frequency sub-band (0-0.5Hz) to reconstruct the signal to reflect the long-term trend. For example, the optical signal path loss of a node increases by 0.3dB per month, corresponding to the accumulation of micro-deformations caused by concrete creep. Select high-frequency sub-bands (2-4Hz) to reconstruct the signal to capture short-term anomalies. For example, a certain equipment start-up and stop caused transient vibration (lasting 5 seconds), resulting in a sudden increase of 20μs² in delay jitter. Align the steady-state and transient components according to the original time window to generate a dual-channel multi-scale feature matrix with the dimension of the number of nodes × the number of time windows × 2 (steady state + transient). The multi-scale features of each node can be described as a long-term attenuation trend (steady state) superimposed on a short-term fluctuation (transient).

[0099] 203. Based on the multi-scale feature matrix, the feature contribution of each time granularity is adjusted by a dynamic weight optimization algorithm to generate an optimized fusion feature matrix;

[0100] In step 203, the feature contribution quantifies the influence of different time scale features on failure prediction, for example, the steady-state component may contribute 70% weight and the transient component contributes 30%.

[0101] In the embodiment of the present application, a reinforcement learning framework is used to achieve weight adaptive optimization. Based on historical data statistics, the initial contribution of the steady-state component to long-term failure prediction is set to 70%, and the contribution of the transient component to sudden failure is set to 30%. Design a reward function, calculate the reward value based on the matching degree between the prediction result and the actual failure event (such as the degree of improvement in accuracy), and iteratively update the weight through the Q-learning algorithm: if the transient component successfully warns of a sudden failure within a certain time window, its contribution will increase by 5% (such as from 30%→35%). The weight of the steady-state component is not less than 50% to avoid false alarms caused by excessive reliance on short-term fluctuations. Multiply the optimized weights element by element with the multi-scale feature matrix to generate an optimized fusion matrix. For example, when a node is in the stage of frequent equipment start-stop, the transient weight is increased to 40%, and the steady-state weight is reduced to 60%.

[0102] 204. Perform state transition analysis on the optimized fusion feature matrix to calculate a piecewise linear function of the historical state attenuation feature of the optical signal transmission path;

[0103] In step 204, the piecewise linear function divides the attenuation process of the optical signal path into multiple stages, and each stage uses a linear slope to describe the state change speed, for example: normal period (slope 0.1dB / day), warning period (slope 0.5dB / day), fault period (slope 2.0dB / day).

[0104] In the embodiment of the present application, a hidden Markov model (HMM) is used to analyze the state transition law, and three types of states are defined according to the numerical distribution of the optimized fusion matrix. In the normal state, the path loss slope is ≤0.2dB / day, and the delay jitter is ≤10μs².

[0105] The slope of the warning state is 0.2-1.0dB / day, and the delay jitter is 10-50μs². The slope of the fault state is ≥1.0dB / day, and the delay jitter is ≥50μs². The state transition frequency of adjacent time windows in the historical data is counted to generate a transition probability matrix. For example: the probability of normal→warning is 20%, and the probability of warning→fault is 30%. For continuous time windows in the same state, calculate the average value of their path loss slope as the linear attenuation rate of this stage. Generate a state transition trajectory for each node, for example: "Node A has gone from normal→warning→fault in the past 30 days, with a slope from 0.1→0.6→1.8dB / day".

[0106] 205. Map the numerical interval of the optimized fusion feature matrix to an initial failure probability based on the piecewise linear function, perform physical constraint verification on the initial failure probability distribution, and output the failure probability of the communication node.

[0107] In step 205, the physical constraint verification combines the actual mechanical parameters of the building structure (such as concrete compressive strength and steel yield strength) to screen the failure probability that meets the actual engineering situation.

[0108] In the embodiment of the present application, the Bayesian probability framework and the physical rule base are used to complete the final verification, and the values ​​of the optimized fusion matrix are mapped to the initial failure probability according to the piecewise linear function. For example: normal state → probability 0.1, warning state → probability 0.5, fault state → probability 0.9. The compressive strength of the area where the node is located is extracted from the building information model (BIM) (such as ≥30MPa as the safety threshold). If the initial probability of a node is 0.8, but its regional compressive strength is 25MPa (lower than the threshold), the probability is reduced to 0.4 (due to physical conditions, it is impossible to reach a high failure risk). The verified probability is normalized to generate a final failure probability distribution diagram. For example: "Node B probability 0.75 (high risk), node C probability 0.35 (medium risk)".

[0109] A large data center park needs to predict the failure risk of core optical communication nodes during the typhoon season. During the typhoon, the strain gradient of the roof steel beam was monitored to increase from 0.1 to 0.3με / mm, and the main vibration frequency increased from 8Hz to 15Hz (close to the structural resonance frequency). The path delay jitter from the core node to the backup center increased from 5μs² to 35μs², and the optical signal-to-noise ratio decreased from 30dB to 22dB. The mechanical weight increased to 58% (due to increased deformation), and the communication weight was 42%. The steady-state component showed that the path loss increased by 0.4dB per month (long-term aging trend). The transient component captured 10 transient vibration events (single lasting 3-10 seconds) during the typhoon, resulting in a delay jitter peak of 50μs². Reinforcement learning increased the weight of the transient component from 30% to 45% and reduced the steady-state weight from 70% to 55% based on multiple sudden failures during the typhoon. The node status changes from normal to warning to fault, and the path loss slope changes from 0.1 to 0.7 to 1.5 dB / day, triggering the warning mechanism. The initial probability is 0.9, but the area where the node is located uses a high-strength steel structure (compressive strength 45 MPa). After physical verification, it is corrected to 0.75 (still high risk), triggering the backup optical cable switch.

[0110] In summary, this solution improves the node failure prediction accuracy to more than 95% and reduces the false alarm rate to less than 5% through multi-modal fusion and dynamic optimization. In typhoon scenarios, a 24-hour warning is achieved before the core node fails, and the backup path switching time is compressed to less than 1 second, ensuring zero interruption of the data center network. Physical constraint verification eliminates the 15% false alarm probability, significantly improving the credibility and engineering applicability of the prediction results.

[0111] In order to construct physically isolated backup optical cable paths and achieve efficient fusion of multi-path signals, firstly, based on the optical cable topology data and node failure probability distribution, an improved path planning algorithm is used to select backup optical cable paths that have no physical overlap with high-failure nodes (such as flying cables outside the building or underground redundant pipelines) to ensure that the number of path hops and transmission delay meet the disaster recovery requirements; then, a spatial diversity receiving system with multiple antenna arrays is deployed to synchronously collect the optical signals of the primary and backup paths, and the phase offset caused by the path difference is eliminated through time-frequency joint analysis methods (such as short-time Fourier transform and dynamic time warping); further, a cross-modal attention mechanism is introduced to align the time domain jitter characteristics and frequency domain energy distribution of the primary and backup signals, and extract the signal components with the strongest complementarity; finally, the multi-path signals are encapsulated into a redundant signal set according to time-space labels, providing a high-availability data source for subsequent shard storage and rapid reconstruction.

[0112] In some embodiments, a backup optical cable path physically isolated from a high failure probability node is extracted from the time-space association relationship, the primary and backup path optical signals are synchronously received through a spatial diversity reception technology, and the time-frequency domain features of the primary and backup path optical signals are cross-modally aligned in combination with the multimodal model to generate a redundant signal set, including:

[0113] 301. Traverse all candidate optical cable paths based on the spatiotemporal association relationship, select candidate optical cable paths that meet the minimum geographical isolation distance with the high failure probability node as backup optical cable paths, and exclude candidate optical cable paths that have overlapping routes with the main optical cable path;

[0114] In step 301, the minimum geographic isolation distance is the physical space distance threshold (e.g., 500 meters) between the backup optical cable path and the high failure probability node, ensuring that both are not damaged at the same time in extreme events (e.g., earthquakes, explosions). Overlapping routes are the overlapped parts of the primary and backup paths in physical routes (e.g., underground pipelines, relay nodes) or geographical areas (e.g., the same building floor, the same cable trench), which may cause common cause failures.

[0115] In the embodiment of the present application, a path planning algorithm and topological analysis technology are used to implement physical isolation screening, and a list of high failure probability nodes (e.g., nodes with a failure probability ≥ 0.8) and an optical cable topology database are input, the latter containing information such as the geographic coordinates, pipeline distribution, and relay node locations of all optical cable paths. Set a geographic isolation distance threshold (e.g., 500 meters) and a route overlap threshold (e.g., the length of the path overlap segment does not exceed 20% of the total path). Traverse all candidate backup paths, calculate the minimum Euclidean distance between each path and the high failure node (based on three-dimensional geographic coordinates), and eliminate paths with a distance less than the threshold. For example, if a candidate path is only 300 meters away from a high-risk node, it is excluded. The overlapping routes of the main and backup paths are analyzed by a path similarity detection algorithm (based on sequence alignment technology): if the two paths share the same underground pipeline for more than 200 meters (assuming that the total length of the main path is 1000 meters), it is determined that the isolation requirements are not met. Filter out backup paths that meet both geographic isolation and route isolation. For example, a path is chosen that bypasses the building perimeter and passes through an independent underground pipeline corridor, which is 600 meters away from the high-risk node and has no pipeline overlap with the main path.

[0116] 302. Deploy diversity receiving devices at the receiving ends of the main optical cable path and the backup optical cable path, and realize synchronous reception of optical signals of the main and backup paths through space diversity receiving technology;

[0117] In step 302, the diversity receiving device is a hardware system consisting of a multi-antenna array, a high-precision clock synchronization module, and a signal preprocessing unit, which is used to synchronously receive the primary and backup path optical signals and eliminate multipath interference. Spatial diversity reception technology uses multiple antennas to receive different copies of the same signal, and uses the spatial independence of the signal to improve the reception quality.

[0118] In the embodiment of the present application, multiple-input multiple-output (MIMO) technology and a precise clock synchronization scheme are used to achieve signal reception. A 4×4 MIMO antenna array is deployed at the receiving end of the main path and the backup path, and the antenna spacing is set to an integer multiple of the wavelength of the optical signal (for example, a 1550nm wavelength corresponds to a spacing of 775nm) to reduce signal correlation. Install a high-precision clock synchronization module (such as a Beidou / GPS dual-mode timing chip) to transmit the clock signal through optical fiber to ensure that the time synchronization error of the receiving end of the main and backup paths is ≤1 nanosecond. After the optical signal of the main and backup paths is converted into a baseband digital signal by photoelectric conversion, it is received by the MIMO array and converted into a baseband digital signal. Adaptive equalization technology (such as the minimum mean square error algorithm) is used to eliminate multipath effects and noise interference, and output a time-aligned IQ signal stream. The synchronized main and backup signals are encapsulated into data packets according to the timestamp, and path labels (such as "main path-node A" and "backup path-node B") are attached.

[0119] 303. Extract time domain waveform features and frequency domain spectrum features from the primary and backup path optical signals after noise suppression respectively, input them into the multimodal model, and use a cross-modal alignment method to fuse the time and frequency domain features of the primary and backup path signals to generate a redundant signal set.

[0120] In step 303, the time domain waveform features reflect the parameters of the optical signal waveform change, such as the signal amplitude mean (measurement of average power), peak-to-average ratio (reflection of signal distortion), and zero crossing rate (characterization of signal stability). The frequency domain spectrum features describe the parameters of the optical signal frequency distribution, such as the main lobe width (reflection of signal bandwidth), side lobe attenuation (measurement of anti-interference ability), and spectrum flatness (characterization of signal quality). Cross-modal alignment dynamically integrates the time domain and frequency domain features of the primary and backup paths to extract the most complementary signal components.

[0121] In the embodiment of the present application, time-frequency analysis and cross-modal attention mechanism are used to realize feature fusion, and a sliding window analysis is performed on the main and standby signals (window length 256 milliseconds, overlap rate 75%), and the peak-to-average ratio of each window is calculated (for example, the peak-to-average ratio of the main path increases from 2.5 to 4.0, indicating that the waveform distortion is aggravated). The time-frequency energy matrix is ​​generated by short-time Fourier transform (STFT), and the main lobe energy ratio is extracted (for example, the main lobe energy ratio of the standby path accounts for 85%, indicating that the spectrum concentration is high). A dual-channel neural network model is constructed to process the time domain and frequency domain features of the main and standby paths respectively. If the main path is severely distorted in the time domain (peak-to-average ratio>3.5), but the standby path frequency domain is stable (main lobe energy ratio>80%), a higher weight (for example, 0.7) is assigned to the standby path frequency domain feature. If the standby path is stable in the time domain (peak-to-average ratio<2.8), its time domain feature weight is increased (for example, 0.6). The aligned main and standby signal features are spliced ​​according to the time window and encapsulated as a redundant signal set. For example, the redundant set of a time window includes: main path time domain features (peak-to-average ratio 3.8, weight 0.3), frequency domain features (main lobe energy 70%, weight 0.2), backup path time domain features (peak-to-average ratio 2.5, weight 0.4), frequency domain features (main lobe energy 85%, weight 0.6).

[0122] A transnational submarine optical cable system needs to build redundant paths for core nodes located in the earthquake zone to ensure communication continuity during earthquake disasters. High-failure node A is located in an active submarine earthquake zone (failure probability 0.9). After traversing the candidate paths, backup path B is selected. Path B detours to the sea area far away from the earthquake fault zone, and the geographical isolation distance from node A is 800 meters. Path B has no shared relay nodes with the main path, and the routing overlap is only 5% (only 50 meters of the landing station entrance section is shared). 4×4 MIMO antenna arrays are deployed at the main path receiving station (land end) and the backup path receiving station (island end), with an antenna spacing of 775nm. Beidou timing signals are transmitted through optical fiber, and the clock synchronization error is controlled at 0.8 nanoseconds to ensure the time alignment of the main and backup signals. During the earthquake, the peak-to-average ratio of the main path signal in the time domain suddenly increased from 2.5 to 4.2 (the waveform was severely distorted), but the main lobe energy of the backup path in the frequency domain remained at 85%. After cross-modal alignment, the frequency domain feature weight of the backup path is increased to 0.75, and the time domain feature weight of the main path is reduced to 0.25, generating a high-confidence redundant signal set. After node A was interrupted by the earthquake, the system switched to the backup path based on the redundant set, with a signal recovery time of ≤50 milliseconds and a packet loss rate reduced from 15% to 0.05%.

[0123] In summary, this solution reduces the risk of common cause failure by 90% through the dual isolation strategy of geography and routing; the spatial diversity reception technology enables the signal synchronization accuracy to reach the sub-nanosecond level, effectively suppressing multipath interference; the redundant signal set generated by cross-modal feature fusion improves the signal-to-noise ratio by more than 12dB in extreme scenarios. The implementation example shows that the system can achieve 50 millisecond-level rapid recovery in extreme events such as earthquakes and wars, and the utilization rate of redundant resources is increased by 40%, which is suitable for scenarios with high reliability requirements such as submarine optical cables and military communications.

[0124] In order to construct a multimodal hash identifier that integrates optical communication and building mechanics features, the data shards are first preprocessed in a standardized manner, and their optical signal wavelength features (such as main wavelength, spectral width) are extracted and associated with the building compressive strength parameters of the area (such as concrete grade, node stiffness). The frequency domain distribution of wavelength features and the spatial mechanical characteristics of compressive parameters are learned through the dual-channel encoding network of the multimodal model, and a cross-modal attention mechanism is designed to dynamically fuse the two types of features and map them to a unified semantic space. A hierarchical hash coding strategy is used to generate a binary hash code containing wavelength-compressive joint information. Finally, the uniqueness and anti-collision of the hash identifier are optimized through the Hamming distance check and collision detection algorithm to form a multimodal hash identifier system that can support rapid retrieval and verification. In some embodiments, the data shards are associated with the optical signal wavelength features and building structure compressive strength parameters of remote nodes through the multimodal model to generate a multimodal hash identifier, including:

[0125] 401. Perform frame and window processing on the optical signal of the remote node based on the multimodal model, and adaptively adjust the frame and window according to the building vibration frequency to construct an optical signal feature sequence.

[0126] In step 401, the frame-by-frame windowing process cuts the continuous optical signal into data segments (frames) of fixed duration, and applies a window function (such as a Hamming window) to each frame signal to suppress spectrum leakage, and the window length is dynamically adjusted according to the building vibration frequency. The optical signal feature sequence contains the time series data of parameters such as the main wavelength (such as 1550.12nm), spectral width (such as 0.03nm), and polarization state (such as 45 degrees) of each frame signal.

[0127] In the embodiment of the present application, adaptive framing technology is used to implement signal processing. Acceleration sensors deployed at key nodes of the building (such as load-bearing beams) are used to monitor the main frequency of structural vibration in real time (such as 4Hz low-frequency vibration detected during a typhoon). Dynamic window length adjustment sets the window length according to the vibration period. For example, 4Hz corresponds to a 250 millisecond window length to ensure that each window covers the complete vibration period and avoids signal truncation and distortion. High-resolution spectral analysis is performed on each frame of the signal to extract the main wavelength (with an accuracy of picometers) and spectral width (reflecting signal stability). The polarization angle of the optical signal is obtained through the polarization state detection module for subsequent anti-interference analysis. The above parameters are arranged in chronological order to form a characteristic sequence of the optical signal, for example, 3600 characteristic points are generated per hour (one window per second).

[0128] 402. Map the compressive strength parameters of the building structure to a three-dimensional space grid coordinate system, calculate the dynamic compressive strength attenuation factor of each grid unit based on the three-dimensional space grid coordinate system, and generate a compressive strength characteristic tensor based on the compressive strength attenuation factor.

[0129] In step 402, the dynamic compressive strength attenuation factor quantifies the degree to which the compressive strength of the building grid unit decays over time under environmental stress (such as temperature and humidity changes, mechanical loads) (such as 0.5% decay per month). The compressive strength characteristic tensor is a collection of compressive strength, material properties (such as concrete grade), attenuation factor and coordinate information of each unit in the three-dimensional spatial grid.

[0130] In the embodiment of the present application, grid modeling and real-time sensor fusion are used to divide the building into cubic grids of 50cm×50cm×50cm. For example, a super high-rise building may contain more than 1 million grid cells. The material properties of the grid cells are extracted from the building information model (BIM) (such as the compressive strength of C60 concrete is 60MPa). A temperature and humidity sensor network is deployed to monitor the environmental parameters of each grid in real time (such as areas with humidity > 80% are marked as high risk). A compressive strength attenuation model is established based on historical data and real-time sensor data. For example, in a high temperature and high humidity environment, the compressive strength of concrete decays by 0.8% per month. A dynamic attenuation factor label is generated for each grid cell (such as a grid G-123 attenuation factor of 0.007 / month). The grid coordinates, compressive strength, and attenuation factor are integrated into a four-dimensional tensor (X×Y×Z×3) to support fast spatial retrieval.

[0131] 403. Establish a mapping relationship table between the optical signal feature sequence and the compressive strength feature tensor, and generate a joint feature matrix including wavelength and compressive strength coupling weights based on the mapping relationship table.

[0132] In step 403, the wavelength and compression resistance coupling weight dynamically adjusts the contribution weight of the optical signal feature according to the degree of compression resistance attenuation (for example, the weight is reduced if the attenuation is fast).

[0133] In the embodiment of the present application, cross-modal spatiotemporal alignment and dynamic weight allocation are used to associate the timestamp of the optical signal feature sequence (such as 2023-10-01 14:00:00) with the spatial coordinates of the compressive strength tensor (such as the east grid of the 80th layer).

[0134] For example, the optical signal characteristics at a certain moment correspond to the compressive resistance parameters of the 80th-81st floor area on the east side of the building. If the compressive strength attenuation rate of a grid unit exceeds the threshold (such as monthly attenuation > 0.6%), the weight of the wavelength characteristics of the optical signal in this area is reduced from 0.7 to 0.4 to reduce dependence on unreliable areas. If the compressive strength is stable (attenuation rate < 0.2%), the weight of the polarization state characteristics of the optical signal is increased to 0.8. The weighted optical signal characteristics (wavelength, spectral width) are concatenated with the compressive resistance parameters (intensity, attenuation factor) to form a joint feature matrix with the dimensions of time × space × 6 (6 feature channels).

[0135] 404. Eliminate correlation between features of the joint feature to generate a multimodal hash identifier with anti-collision characteristics.

[0136] In step 404, the anti-collision feature ensures the uniqueness of hash identifiers of different data shards through a coding strategy, and the probability of collision is less than 0.001%.

[0137] In an embodiment of the present application, hierarchical hash coding and redundancy check are used to perform principal component analysis (PCA) on the joint feature matrix, extract the top three principal components with the highest variance contribution (covering 96% of the information), and eliminate redundant features. The first layer of coding maps the principal components to 128-bit binary codes, for example, by discretizing the principal component values ​​into binary bits through segmented quantization. The second layer of verification adds an 8-bit cyclic redundancy check code (CRC) to detect and correct bit errors in transmission. If the Hamming distance between the two slice hash codes is ≤3 (that is, the difference does not exceed 3 bits), the recoding mechanism is triggered to regenerate a unique hash code by introducing a compressive strength priority label (such as higher intensity areas first).

[0138] The optical communication system of a cross-sea bridge needs to ensure the recoverability of data fragmentation in a salt spray corrosion environment. The vibration frequency of the bridge main beam is monitored to be 2Hz (caused by wave impact), and the window length is set to 500 milliseconds. The main wavelength of the optical signal extracted by frame is 1550.15nm (salt spray causes wavelength drift of ±0.05nm), and the spectrum width is 0.05nm. The grid of the bridge cable anchorage area (compressive strength 550MPa) has a monthly attenuation rate of 1.2% due to salt spray corrosion. Generate a compressive tensor and mark the grid parameters as [550MPa, 1.2%, X=200, Y=50, Z=30]. Due to the high compressive attenuation rate, the optical signal weight in this area is reduced from 0.8 to 0.3, and the weight of the adjacent low attenuation area is increased to 0.7. The joint matrix records the coupling characteristics of the optical signal wavelength 1550.15nm and the compressive strength 550MPa at this moment. PCA retains three principal components, encoded as a 128-bit hash code "0x5F3A...", and an additional CRC-8 checksum "0x2C". The final hash identifier "0x5F3A...2C" is stored in the blockchain distributed ledger. After a node fails due to corrosion caused by salt spray, the system matches the same wavelength shard in the high pressure resistance area (550MPa) according to the hash identifier, and the data recovery time is ≤80 milliseconds.

[0139] In summary, this solution achieves accurate fusion of optical signals and building mechanical parameters (matching accuracy ≥ 98%) through adaptive framing and compression attenuation modeling; dynamic weight allocation and hierarchical hash coding technology make the hash identification anti-collision rate > 99.99% and the mismatch rate less than 0.05%. In the cross-sea bridge scenario, the data recovery time is shortened to 80 milliseconds, and the redundant resource consumption is reduced by 35%. It is suitable for high-risk environments such as salt spray corrosion and earthquake shock, and significantly improves the data reliability of key infrastructure.

[0140] In order to improve the self-healing ability and data reliability of the communication network, firstly, based on the wavelength characteristics of the optical signal in the multimodal hash identifier (such as the main wavelength tolerance ±0.1nm) and the compressive strength threshold (such as ≥30MPa), parallel hash matching is performed in the distributed node cluster to screen out the data shard set that meets the wavelength consistency and is located in the high-compressive strength grid; then, the time-frequency complementary characteristics of the multipath signal in the redundant signal set are extracted (such as the main path time domain distortion but the backup path frequency domain stability), and the path difference interference is eliminated through dynamic phase alignment and weighted fusion technology; finally, the reconstruction results are physically verified in combination with the real-time mechanical parameters of the building structure (such as node stiffness, crack extension threshold), and abnormal data that does not meet the material strength constraints is eliminated, and a high-integrity key information stream is output. In some embodiments, the data shards with the same optical signal wavelength characteristics and located in the high-compressive strength area are matched from the remote node according to the multimodal hash identifier, and the key information is reconstructed based on the complementary characteristics of the multipath signals in the redundant signal set, including:

[0141] 501. Establish a distributed hash mapping table in a remote node according to the multimodal hash identifier, perform a neighbor search on the hash mapping table, and screen and match data fragments having the same optical signal wavelength characteristics based on the neighbor search;

[0142] In step 501, a distributed hash map is an index table stored in a remote node, which records the unique hash identifier (such as a 64-bit binary code) of the data shard, the wavelength characteristics of the optical signal (such as the main wavelength, the spectrum width), the storage location coordinates and other information, for fast retrieval of the target shard. The nearest neighbor search searches for shards with high similarity to the target hash code in the Hamming space of the hash identifier, and the similarity is measured by the Hamming distance (the number of difference bits), for example, a Hamming distance ≤ 3 indicates high similarity.

[0143] In an embodiment of the present application, a distributed index architecture and local sensitive hashing technology are used to achieve efficient matching, and the hash identifiers of data shards are distributed and stored in different nodes according to the consistent hashing algorithm. Each node manages a specific range of hash codes (such as node A is responsible for hash codes 0x0000~0x3FFF). An optical signal wavelength label (accuracy ±0.1nm) and the three-dimensional coordinates of the storage location are attached to each shard (such as node B is located on the east side of the 80th floor of the building). Enter the target hash identifier (for example, 0x3A7D) and query the shards with a Hamming distance ≤3 in parallel in the distributed cluster. Perform a secondary wavelength verification on the candidate shards: if the difference between its main wavelength and the target wavelength is ≤0.1nm (such as target 1550.12nm, candidate 1550.18nm), it is marked as a valid match. Return the physical addresses of all qualified shards (such as node C stores shard S-15).

[0144] 502. In the data shard matching process, the pressure sensor data of the remote node is collected in real time, a node compressive strength dynamic matrix is ​​constructed, the compressive coefficient of each node is calculated through the node compressive strength dynamic matrix, nodes with large compressive coefficients are retained, and data shards located in the high compressive strength area are screened out based on the nodes with large compressive coefficients;

[0145] In step 502, the node compressive strength dynamic matrix is ​​a two-dimensional matrix updated in real time, recording the compressive strength (unit: MPa), environmental stress (such as temperature, humidity), historical load data and other parameters of the node area. The compressive coefficient is a scoring index (0-1 range) that combines the compressive strength and environmental stress, and is used to quantify the reliability of the node area.

[0146] In the embodiment of the present application, a dynamic weighted scoring model and real-time sensor fusion are used to achieve pressure resistance screening. The pressure sensor network (such as fiber Bragg grating sensor) deployed in the node area collects the pressure resistance in real time (for example, the pressure resistance in the area where node D is located is 50MPa). The temperature and humidity sensors monitor environmental parameters (such as temperature 35°C and humidity 80%) and calculate the environmental stress score (for example, the high temperature and high humidity score is 0.8, indicating high risk). The weight of the pressure resistance is 60%, and the weight of the environmental stress is 40%. For example: the pressure resistance is 50MPa (standardized to 0.9) and the environmental stress score is 0.8 (high risk, standardized to 0.2). The pressure resistance coefficient = 0.6×0.9 + 0.4×(1-0.8) = 0.54 + 0.08 = 0.62 sets a threshold (such as ≥0.7) to screen high-reliability nodes, and the shards below the threshold are eliminated. Only the data shards in the nodes with the pressure resistance coefficient that meets the standard (such as the node E coefficient 0.75) are retained to ensure that the data is stored in the high pressure resistance area.

[0147] 503. Establish a space-frequency joint dictionary according to the complementary characteristics of the multipath signals in the redundant signal set, output signal reconstruction error minimization and energy convergence conditions based on the space-frequency joint dictionary, and output key information after convergence verification based on the signal reconstruction error minimization and energy convergence conditions.

[0148] In step 503, the space-frequency joint dictionary contains a library of complementary feature templates of multipath signals in the time domain (such as signal waveform, phase difference) and frequency domain (such as spectrum energy distribution, main lobe width) to guide signal reconstruction. During the energy convergence condition signal reconstruction process, the reconstruction error is required to be gradually reduced and the total energy fluctuation is stable (such as fluctuation amplitude ≤ 5%) to ensure that the output result is reliable.

[0149] In the embodiment of the present application, iterative optimization and dynamic weight allocation are used to achieve reconstruction, and time-frequency analysis is performed on the primary and backup path signals in the redundant signal set. The time domain waveform is stable but the frequency band energy is low (such as high-frequency attenuation due to path aging), and the frequency band energy is concentrated but there is phase jitter in the time domain (such as due to external interference). The complementary features of the two types of signals are extracted (such as the time domain stability of the main path and the frequency domain energy of the backup path), and dictionary entries are constructed. The entry with the highest time domain stability (such as the main path weight 0.7) is selected from the dictionary to generate the initial reconstructed signal. The mean square error between the reconstructed signal and the original signal is calculated (such as the initial error 20%). If the error does not converge, increase the frequency domain weight of the backup path (such as from 0.3 to 0.5), regenerate the signal and evaluate the error. When the error decreases for three consecutive iterations and the energy fluctuation is ≤5% (for example, the error is from 20%→15%→10%→8%), convergence is determined and the final reconstructed data is output.

[0150] During the typhoon, a submarine optical cable system needs to restore critical data that failed due to node corrosion. The target hash identifier is 0x5F3A (wavelength 1550.15nm), and the neighbor search matches the fragment S-25 (Hamming distance 2, wavelength difference 0.07nm), which is stored in node F (submarine relay station). The compressive strength of the area where node F is located is 55MPa (high-pressure steel armor protection), the temperature is 10℃, the humidity is 30%, and the compressive coefficient is calculated to be 0.82 (higher than the threshold 0.7), which is confirmed as a high-pressure area. In the redundant signal set, the main path is completely lost due to cable breakage, and the backup path signal has severe time domain jitter (peak-to-average ratio 4.5) but concentrated frequency domain energy (main lobe energy 90%). The frequency domain features (weight 0.8) of the backup path in the space-frequency dictionary are selected for reconstruction. After 4 iterations, the error is reduced from 25% to 6%, the energy fluctuation is 3%, and the complete data packet is output. The reconstructed data passes the cyclic redundancy check (CRC-32), the packet loss rate is 0.01%, the recovery time is 150 milliseconds, and the service switching is imperceptible.

[0151] In summary, this solution achieves a data shard matching accuracy of more than 99% through hash neighbor retrieval and dynamic screening of compression coefficients; multipath signal complementary reconstruction technology compresses the key information recovery time to within 150 milliseconds, with an error rate of less than 5%. In the submarine optical cable scenario, data shards damaged by extreme weather were successfully restored, and redundant resource consumption was reduced by 50%, which is suitable for the real-time recovery needs of key data in high-reliability communication networks (such as military and aerospace).

[0152] In order to solve the phase aliasing and energy diffusion problems caused by propagation path differences in multipath signals, the phase ambiguity between multipath signals is first eliminated by space-frequency joint projection, and a multipath complementary feature tensor containing path differences and frequency domain attenuation weights is constructed. The space-frequency joint dictionary matrix is ​​generated by decomposing the tensor, and the incoherence of the basis function is optimized through orthogonal constraints to ensure that the dictionary atoms have sparse representation capabilities for multipath interference; then a signal reconstruction error minimization model is designed, the path energy weighting factor and phase calibration coefficient constraints are introduced, the error function is optimized in combination with the gradient descent strategy, and the dynamic relaxation factor is iteratively adjusted in conjunction with the energy conservation equation to balance the local residual and the global convergence speed; finally, the time-frequency energy distribution of the reconstructed signal is calculated, the residual standard deviation and the energy flux conservation rate are calculated, and a double verification mechanism is used to screen the stable convergence path, and the key signal feature vector for anti-interference is output.

[0153] In some embodiments, in step 503, establishing a space-frequency joint dictionary according to the complementary characteristics of the multipath signals in the redundant signal set, outputting signal reconstruction error minimization and energy convergence conditions based on the space-frequency joint dictionary, and outputting key information after convergence verification based on the signal reconstruction error minimization and energy convergence conditions include:

[0154] 5031. Extract the spatiotemporal propagation characteristics of the complementary features of the multipath signals from the redundant signal set, eliminate the phase ambiguity of the sensor array through space-frequency joint projection, and output the standardized multipath complementary feature tensor;

[0155] In step 5031, the complementary characteristics of the multipath signal include the time delay difference (time delay of different propagation paths), Doppler shift (frequency offset caused by relative motion), polarization response (change in polarization direction of the signal after reflection) and other time-space propagation characteristics of the multipath signal. Phase ambiguity is the signal direction estimation error caused by the phase superposition of multipath signals in the sensor array. The standardized multipath complementary characteristic tensor is a three-dimensional data matrix formed by aligning the time-frequency characteristics of the multipath signal through space-frequency joint projection, which contains parameters such as path difference weight and frequency domain attenuation coefficient.

[0156] In the embodiment of the present application, firstly, the original signal in the redundant signal set is subjected to time-frequency analysis to extract the delay difference (calculated by the cross-correlation function), Doppler frequency shift (based on short-time Fourier transform peak detection) and polarization response (using the covariance matrix decomposition of the polarization sensitive array) of each path signal. Subsequently, the phase ambiguity of the sensor array is eliminated (such as the wave direction estimation correction based on the MUSIC algorithm) through the space-frequency joint projection technology (mapping the signal to the joint space of spatial domain beamforming and frequency domain subband decomposition). Finally, the time-frequency characteristics of the multipath signal are normalized, and the standardized multipath complementary feature tensor is output.

[0157] 5032. Optimizing the incoherence between basis functions of the multi-path complementary feature tensors through orthogonal constraints to generate a space-frequency joint dictionary with path difference compensation characteristics;

[0158] In step 5032, the orthogonal constraint requires that the basis functions (spatial basis functions and frequency domain atomic nuclei) in the dictionary be orthogonal in the vector space to reduce redundancy. The path difference compensation feature is that the dictionary atoms can adaptively match the signal characteristics of different propagation paths and compensate for the delay and attenuation differences between paths.

[0159] In the embodiment of the present application, the tensor outputted from step S5031 is decomposed into spatial basis functions (describing the spatial propagation direction of the signal) and frequency domain atomic nuclei (describing the frequency domain energy distribution of the signal), and the orthogonal matching pursuit algorithm (OMP) is used to optimize the orthogonality between the basis functions to ensure the sparse representation capability of the dictionary atoms for multipath signals. Specifically, the singular value decomposition (SVD) is used to extract the principal components as the initial basis functions, and then the incoherence of the basis functions is adjusted by the iterative reweighted least squares method (IRLS), and finally a space-frequency joint dictionary matrix capable of distinguishing signals of different paths is generated.

[0160] 5033. Construct the reconstruction error objective function based on the space-frequency joint dictionary, optimize the second-order derivative matrix of the error function through the gradient descent strategy, and output the signal reconstruction error minimization and energy convergence conditions;

[0161] Reconstruction error objective function: measures the reconstruction error of the original signal and the linear combination of dictionary atoms, and the goal is to minimize the L2 norm of the error.

[0162] In step 5033, the second-order derivative matrix optimization analyzes the curvature characteristics of the objective function through the Hessian matrix to accelerate the convergence of gradient descent.

[0163] In the embodiment of the present application, a reconstruction error function (such as a least squares loss function) is constructed based on a space-frequency joint dictionary, and a path energy weighting factor (calculated by the frequency domain attenuation coefficient of step S5031) and a phase calibration coefficient (corrected by the polarization response error) are added. The quasi-Newton method (such as the L-BFGS algorithm) is used to calculate the gradient and Hessian matrix of the objective function, and the iteration step size is dynamically adjusted to avoid local minima. At the same time, an energy convergence condition (such as a threshold value of the ratio of the total signal energy to the residual energy) is introduced, and the constraints are integrated into the optimization process through the Lagrange multiplier method, and finally the joint optimization parameters that satisfy error minimization and energy conservation are output.

[0164] 5034. Based on the signal reconstruction error minimization and energy convergence conditions, a double verification mechanism is used to screen the stable convergence path and output key information to resist multipath interference.

[0165] In step 5034, the double verification mechanism includes residual standard deviation threshold determination (eliminating abnormal paths) and energy flux conservation verification (verifying signal energy closure). The path signal that passes the verification of the stable convergence path has delay consistency, energy traceability and phase stability.

[0166] In the embodiment of the present application, the time-frequency energy distribution of the multipath signal is reconstructed using the optimized parameters of step S5033, and the residual standard deviation (based on the difference statistics between the reconstructed signal and the original signal) and the energy flux conservation rate (the ratio of the signal input energy to the reconstructed energy) of each path are calculated. By setting a dynamic threshold (such as the 3σ principle), the paths with residuals below the threshold are screened, and their energy closure (such as flux error <5%) is verified. Finally, the verified path signals are fused (such as weighted average or maximum likelihood merging), and key anti-interference information (such as direct wave signal parameters) is output.

[0167] In the indoor positioning scenario of drones, multiple reflection surfaces result in 5 propagation paths (direct wave and 4 reflected waves) for the received signal. The delay difference (0~50ns), Doppler frequency shift (-200Hz~+200Hz) and polarization angle offset (0°~30°) of each path are extracted to generate a standardized tensor containing path weights (direct wave weight 0.8, reflected wave weight 0.05~0.1). The tensor is decomposed to generate 10 spatial basis functions and 8 frequency domain atomic nuclei. The dictionary is constructed through orthogonal constraints, and the atomic nucleus energy corresponding to the direct wave accounts for 85%. After optimization, the reconstruction error of the direct wave is reduced to 1.2% (the original error is 15%), and the reflected wave energy converges to less than 5% of the total energy. The direct wave paths with residual standard deviation <0.1 and energy flux error <3% are screened out, and high-precision TOA (arrival time) and carrier phase information are output, and the positioning error is reduced from 1.5m to 0.2m.

[0168] This solution effectively separates the direct wave and reflected wave components in multipath signals through the construction of space-frequency joint dictionary and error-energy joint optimization, suppresses delay spread and phase aliasing interference; uses a double verification mechanism to screen stable paths and improve the reliability of signal reconstruction and positioning accuracy. In complex electromagnetic environments (such as indoors and urban canyons), sub-meter positioning accuracy can be achieved, and the robustness is more than 3 times higher than that of traditional methods.

[0169] In order to improve the timeliness and interpretability of failure prediction, the optimized fusion feature matrix is ​​first segmented into time series, and the path state (such as normal, warning, and fault) is dynamically divided based on the hidden Markov model (HMM). The state transition law is extracted through the sliding window analysis of historical attenuation data; then the piecewise linear regression algorithm is used to fit the attenuation slope in each state interval, and the segmentation boundary is dynamically adjusted in combination with the time-frequency characteristics of the optical signal (such as frequency domain energy entropy and delay jitter variance), and a piecewise linear function matching the physical constraints (such as material fatigue threshold and environmental stress) is generated to quantify the attenuation rate and state continuity at different stages, and provide an interpretable attenuation trajectory model for failure probability mapping. In some embodiments, state transition analysis is performed on the optimized fusion feature matrix, and the piecewise linear function of the historical state attenuation characteristics of the optical signal transmission path is calculated, including:

[0170] 601. Based on the position parameters in the fusion feature matrix, extract the state features in the local neighborhood through a preset neighborhood set to generate a neighborhood state feature set;

[0171] In step 601, the location parameters are fused with the three-dimensional spatial coordinates of the nodes recorded in the feature matrix (e.g., the east side of the 80th floor of the building, coordinates X=120m, Y=80m, Z=30m) and the associated regional labels (e.g., "high-voltage cable well area"). The neighborhood set is centered on the target node and covers all the nodes in its physical neighborhood, which is used to capture the state relevance within a local range (e.g., all nodes within a radius of 50 meters).

[0172] In the embodiment of the present application, spatial correlation analysis and dynamic neighborhood construction are used to realize feature extraction, and the neighborhood radius is dynamically adjusted according to the building structure characteristics (such as floor height, wall isolation) and optical signal propagation characteristics (such as attenuation rate). For example, a radius of 50 meters is set on an open floor (such as a factory building), and it is reduced to 20 meters in a dense cable well area. The neighborhood nodes are screened by a topological analysis algorithm. For example, the neighborhood of node A includes nodes B and C in the same cable duct, and node D on the upper and lower floors. The fusion features of the nodes in the neighborhood are aggregated and analyzed, and the average value (such as 8 microseconds squared) and maximum value (such as 15 microseconds squared) of the delay jitter of the nodes in the neighborhood are calculated to reflect the local signal stability. The distribution range of the main lobe energy of the nodes in the neighborhood frequency domain (such as 85%-92%) is counted to measure the signal spectrum concentration. The aggregation results (such as the delay mean 8μs² and the main lobe energy range 85%-92%) are encapsulated as a neighborhood state feature vector. The neighborhood state features are recorded in chronological order (such as one sample per hour) to form a time series set. For example, a continuous 24-hour feature set can detect the impact of diurnal load variations on attenuation.

[0173] 602. Perform time-varying weight allocation on the neighborhood state feature set, introduce an attenuation coefficient related to the time step for each neighborhood state, and dynamically adjust the nonlinear attenuation gradient of the attenuation coefficient through a piecewise linear interpolation algorithm;

[0174] In step 602, the time-varying weight dynamically adjusts the contribution of the neighborhood state according to the timeliness of the data, for example, new data (nearly 1 week) has a high weight and old data (more than 1 month) has a low weight. The decay coefficient quantifies the degree of reliability decay of the neighborhood state over time, for example, the monthly decay coefficient is 0.95 (i.e., the weight decreases by 5% every month).

[0175] In the embodiment of the present application, the time decay model and adaptive gradient adjustment are used to realize the dynamic weight allocation of the neighborhood state feature set: the latest data (collected on the same day) is assigned an initial weight of 1.0, and the historical data decays according to the time step. For example, the weight of last week's data is 0.9, and the weight of last month's data is 0.7. According to the material aging rate and environmental stress (such as temperature and humidity), the time interval is divided. In the short-term stage (0-3 months), the optical cable material ages slowly, and the weight decays linearly (decreases by 0.05 per month). In the medium-term stage (3-6 months), material fatigue accumulates, and the weight decays rapidly (decreases by 0.1 per month). In the long-term stage (more than 6 months), aging is significant, and the weight decays exponentially (decreases by 0.15 per month). For example, the weight of the data in the 4th month is reduced from 0.85 to 0.75 (a decrease of 0.1), while the weight of the data in the 7th month is reduced from 0.6 to 0.45 (a decrease of 0.15). The attenuated weight is multiplied by the neighborhood feature vector to generate a weighted feature set. For example, the delay jitter characteristic (8μs²) of a neighborhood becomes 8×0.85=6.8μs² after attenuation.

[0176] 603. Perform state transition analysis on the neighborhood state feature set after the attenuation coefficient is adjusted, divide the historical state into multiple intervals, calculate the boundary value jump parameters across the intervals, and generate a piecewise linear function of the attenuation characteristics of the historical state of the optical signal transmission path.

[0177] In step 603, the boundary value jump parameter state interval switching is a sudden change indicator of the attenuation rate, for example, the slope from the normal stage to the warning stage suddenly increases by 3 times. The piecewise linear function divides the historical attenuation process into multiple linear stages, each of which describes the attenuation rate under different states (such as 0.1dB / month in the normal period and 0.5dB / month in the fault period).

[0178] In the embodiment of the present application, state clustering and boundary detection technology are used to realize segmented modeling. Based on the weighted neighborhood feature set, a clustering algorithm (such as K-means) is used to divide historical data into three categories. The normal state delay jitter is ≤10μs², and the main lobe energy is ≥85%. The warning state delay jitter is 10-20μs², and the main lobe energy is 80%-85%. The fault state delay jitter is ≥20μs², and the main lobe energy is ≤80%. For example, a node is in a normal state in 0-4 months, enters the warning state in 5-7 months, and enters the fault state after 8 months. Detect the state switching point (such as jumping from normal to warning in the 5th month), and calculate the change multiple of the attenuation slope before and after the jump. For example, the slope in the normal stage is 0.08dB / month, and the slope in the warning stage rises to 0.32dB / month (an increase of 4 times). The piecewise function fitting performs linear fitting on the data points in each state interval to generate a piecewise attenuation function. The slope in the normal stage (0-4 months) is 0.08dB / month, reflecting slow aging. The slope in the early warning stage (May-July) is 0.32 dB / month, indicating accelerated degradation. The slope in the failure stage (August-December) is 0.68 dB / month, indicating imminent failure.

[0179] A data center park needs to predict the attenuation trend of the core optical cable path to prevent large-scale communication interruptions. The core node is located on the 5th floor of Building 3 in the park (coordinates X=200, Y=150, Z=5), and the neighborhood includes 6 nodes in the same optical cable well (radius 25 meters). The neighborhood delay jitter mean 9μs² and the main lobe energy range 87%-90% are extracted to generate feature vectors [9, 87-90]. The optical cable material is low-smoke halogen-free type with strong aging resistance. The attenuation coefficient is set, and the weight from 0 to 4 months decreases linearly from 1.0 to 0.8 (decrease by 0.05 per month). From April to August, the weight decreases rapidly from 0.8 to 0.6 (decrease by 0.1 per month). After August, the weight decreases exponentially from 0.6 to 0.3 (decrease by 0.15 per month). The weight of the data in the sixth month is 0.7, which becomes 0.7×0.6=0.42 after adjustment. Clustering status: normal (0-5 months, slope 0.07dB / month), warning (6-8 months, slope 0.29dB / month), fault (9-12 months, slope 0.73dB / month). In the 6th month, the slope suddenly increased by 4.14 times (0.07→0.29), triggering the warning signal. According to the piecewise function, the system automatically switched to the backup path at the beginning of the 8th month to avoid business interruption caused by core node failure.

[0180] In summary, this solution improves the attenuation trend prediction accuracy to 97% through dynamic aggregation of neighborhood states and time attenuation weight allocation; the piecewise linear function combined with jump parameter detection can warn of potential failures 4-6 months in advance. In the data center scenario, the core path maintenance response time is shortened to within 48 hours, and the fault false alarm rate is less than 2%, which is suitable for scenarios with strict stability requirements such as high-density optical cable networks and industrial Internet of Things.

[0181] In order to improve the accuracy and robustness of the transmission path state modeling, the dynamic deformation parameters of the building structure (such as strain gradient, vibration energy spectrum) are first time-series segmented and feature extracted, and the spatiotemporal correlation of the optical signal transmission path (such as delay jitter, path loss) is analyzed. The evolution law of mechanical deformation and the spatiotemporal distribution characteristics of optical communication parameters are learned through the dual-channel encoding network of the multimodal model respectively; a cross-modal attention mechanism is designed to dynamically evaluate the contribution of the two types of features to the stability of the transmission path, and the weight ratio (such as deformation feature weight 40%-60%) is adaptively adjusted based on real-time environmental stress (such as temperature and load changes) and historical degradation trends to eliminate scale differences between modes; finally, the weighted feature vectors are aligned according to the spatiotemporal labels to generate a fusion feature matrix, which provides a unified representation space for subsequent state analysis and failure prediction. In some embodiments, the dynamic deformation parameter sequence and the spatiotemporal correlation are fused through a multimodal model, and the weight ratio of the two types of features in the optical signal transmission path is dynamically allocated to generate a fusion feature matrix, including:

[0182] 701. Slice the dynamic deformation parameter sequence into time periods through a sliding window based on the time dimension using a multimodal model, normalize the data in the sliding window, and ensure alignment of the time periods with the spatiotemporal correlation relationship;

[0183] In step 701, the dynamic deformation parameter sequence is the time series data of building structure mechanical parameters collected in real time by a high-density sensor network, such as the strain gradient (unit: microstrain / mm), vibration main frequency (Hz), and load distribution (kN / m2) of the beam-column node. The spatiotemporal correlation relationship describes the spatiotemporal dynamic parameters of the optical signal transmission path, including the number of hops between nodes, transmission delay (microsecond level), path loss (dB / km), etc. The sliding window cuts the continuous time series data into multiple time segments according to a fixed time length (such as 60 seconds), and the window step length (such as 10 seconds) determines the overlap ratio between segments (such as 83%).

[0184] In the embodiment of the present application, dynamic window adjustment and standardization are used to achieve data alignment and unified scale, and the window length is set according to the optical signal sampling frequency (such as 1000 times per second). For example, a 60-second window contains 60,000 sampling points to ensure that the complete building vibration cycle is covered (such as 12 cycles for 5Hz vibration). The dynamic deformation parameters (such as strain gradients) and optical signal parameters (such as delay jitter) are cut synchronously, and the timestamp alignment is ensured by a high-precision clock (such as the GPS-PTP protocol), with an error of no more than 1 microsecond. The data in each window is standardized. For example, the strain gradient value is converted by the Z-score method: subtract the mean and divide by the standard deviation to make the data distribution mean 0 and the variance 1. Optical signal parameters (such as delay jitter) are mapped to the 0-1 interval, for example, the maximum value of the delay jitter 50 microseconds squared corresponds to a normalized value of 1.0. Check the time synchronization of the dynamic deformation and optical signal parameters in the window. For example, a vibration peak moment (such as the 30th second) must be strictly aligned with the corresponding delay jitter surge moment. If the deviation exceeds a threshold (such as ±5 milliseconds), data re-collection is triggered.

[0185] 702. Perform time-step-by-time similarity measurement on the local features of the dynamic deformation parameter sequence and the global features of the spatiotemporal correlation relationship to generate a weight coefficient matrix;

[0186] In step 702, local features are statistical characteristics of dynamic deformation parameters within a single time window, such as strain gradient variance (reflecting deformation fluctuation intensity) and vibration energy peak (characterizing instantaneous impact intensity). Global features are trending optical signal parameters across time periods, such as the quarterly growth rate of path loss (dB / month) and the monthly average change of delay jitter (microseconds squared / month).

[0187] In the embodiment of the present application, trend matching and correlation analysis are used to implement similarity measurement, and the variance of the strain gradient in the window (such as 0.05 microstrain / mm square) and the maximum value of the vibration energy (such as 0.8 joules / Hz) are calculated. The global feature extraction statistics the quarterly growth trend of the optical signal path loss (such as a monthly increase of 0.3 decibels) and the monthly average fluctuation range of the delay jitter (such as 10-30 microseconds squared). Analyze the synchronization of local deformation characteristics (such as strain gradient variance) and global optical signal trends (such as path loss growth). For example, if the strain gradient variance of a window increases by 20% and the path loss growth accelerates, the matching degree is 0.85 (range 0-1). Calculate the Pearson correlation coefficient (such as 0.72) of the vibration energy peak and the delay jitter fluctuation to reflect the linkage between the two in the time dimension. Generate a weight coefficient matrix based on the matching degree and correlation. For example, if the trend matching degree is 0.85 and the correlation is 0.72, the weight coefficient is 0.75 (range 0-1).

[0188] 703. Dynamically allocate weight proportions of two types of features in the optical signal transmission path according to the weight coefficient matrix;

[0189] In step 703, the weight ratio is the contribution ratio of the dynamic deformation parameter and the light signal parameter in the fusion, for example, the deformation weight is 60% and the light signal weight is 40%, which is adjusted in real time according to environmental factors.

[0190] In the embodiment of the present application, environmental adaptive rules and expert experience database are used to realize dynamic adjustment of weights, and real-time data such as temperature (such as 40°C), humidity (such as 85%), and mechanical load (such as 500 kN) are collected. Optical signals are easily affected by thermal expansion and contraction of optical fibers, so the deformation weight is increased to 65% and the optical signal weight is reduced to 35%. Deformation parameters (such as strain gradients) are more sensitive to load changes, and the weight is increased to 70%. The default weight ratio is (60% deformation, 40% optical signal). Combine the similarity measurement results with environmental parameters to generate a weight matrix for time step t. For example, the weight of a window in a high temperature environment is [0.65, 0.35].

[0191] 704. Through matrix dot multiplication and weighted summation operations, the feature matrix of the dynamic deformation parameter sequence and the feature matrix of the spatiotemporal correlation relationship are fused according to the weight proportions to generate a fused feature matrix.

[0192] In step 704, matrix dot multiplication multiplies the weight matrix and the original feature matrix element by element to achieve feature weighting. Weighted summation concatenates or superimposes weighted feature matrices of different modalities according to the channel dimension to generate a fusion matrix.

[0193] In the embodiment of the present application, multi-channel fusion and dimensionality reduction compression are used to achieve efficient characterization. The deformation feature matrix (dimension: time step × number of windows × number of features, such as 100×60×5) is multiplied element by element with the weight matrix (100×60×1) to obtain a weighted deformation matrix. The optical signal feature matrix (100×60×3) is multiplied with the weight matrix to generate a weighted optical signal matrix. The weighted deformation matrix (100×60×5) and the optical signal matrix (100×60×3) are spliced ​​along the feature dimension to form a fusion matrix of 100×60×8. The feature dimension is reduced from 8 to 3 through principal component analysis (PCA), retaining 95% of the information, and finally generating a fusion matrix of 100×60×3. The fusion matrix is ​​losslessly compressed (such as the LZ77 algorithm), and the storage space occupancy is reduced by 60%, supporting real-time transmission and fast retrieval.

[0194] The optical communication system of a cross-sea bridge needs to integrate bridge vibration data and optical signal parameters to predict the stability of the critical path. The vibration data of the main span of the bridge (main frequency of vibration 2Hz, strain gradient 0.1 microstrain / mm) and the optical signal delay jitter (mean 15 microseconds squared) are divided into 60-second windows with a step size of 10 seconds. After normalization, the vibration data distribution is [-1.2, 2.3], and the delay jitter is mapped to [0.3, 0.7]. The vibration variance of a certain window is 0.12 joules / Hz squared, and the global feature is that the optical signal path loss increases by 0.5 dB / month quarterly. The trend matching degree is 0.88, the correlation is 0.75, and the generated weight coefficient is 0.8. The humidity in the sea area where the bridge is located is 90%, triggering the weight adjustment rule, and the deformation weight is increased to 68%, and the optical signal weight is 32%. The vibration matrix (100×60×4) and the optical signal matrix (100×60×2) are weighted and concatenated into 100×60×6, and the fusion matrix 100×60×3 is generated after dimensionality reduction. The fusion matrix is ​​input into the prediction model, and the path instability is warned 3 days in advance. The system automatically switches to the backup submarine optical cable to avoid communication interruption.

[0195] In summary, this solution improves the accuracy of multimodal feature fusion to 94% through dynamic window alignment and adaptive weight allocation; in the cross-sea bridge scenario, it achieves early warning 3 days before the failure of the critical path, with a false alarm rate of less than 4% and a resource utilization rate of 40%. It is suitable for high-reliability communication network operation and maintenance in complex environments such as bridges, tunnels, and data centers, significantly reducing the risk of sudden failures.

[0196] In order to improve the credibility and operability of the prediction results, firstly, based on the attenuation rate thresholds corresponding to each state interval (such as normal, warning, and fault) defined by the piecewise linear function, the optimized fusion feature matrix values ​​(such as time domain jitter variance and frequency domain energy entropy) are mapped to the initial failure probability (such as the range of 0.1-0.9); then, a physical constraint rule library is constructed in combination with the real-time physical parameters of the building structure (such as the compressive strength of the node material, the ambient temperature and humidity, and the crack extension threshold), and the initial probability is corrected: if the compressive strength of the node area is lower than the safety threshold (such as concrete <30MPa) or the environmental stress exceeds the standard (such as humidity >85%), the failure probability weight is reduced according to the rules; finally, the multi-source constraints are fused through the Bayesian probability framework to output a node failure probability distribution map that conforms to the actual engineering situation, ensuring that the prediction results are both data-driven and physically interpretable.

[0197] In some embodiments, based on the piecewise linear function, the numerical interval of the optimized fusion feature matrix is ​​mapped to the initial failure probability, the initial failure probability distribution is physically constrained and verified, and the failure probability of the communication node is output, including:

[0198] 801. Divide the numerical interval of the fusion feature matrix into multiple sub-intervals based on the piecewise linear function, perform linear transformation on the numerical values ​​in the sub-intervals, and generate an initial failure probability;

[0199] In step 801, the piecewise linear function divides the attenuation history of the optical signal path into multiple stages, each stage corresponding to a specific attenuation rate range. For example, the normal stage (low attenuation rate), the warning stage (medium attenuation rate), and the fault stage (high attenuation rate). The initial failure probability is mapped to a probability value between 0 and 1 according to the attenuation stage of the value of the fused feature matrix. The higher the value, the greater the risk of failure.

[0200] In the embodiment of the present application, interval mapping and linear interpolation are used to realize probability generation. According to historical data and engineering experience, the attenuation process is divided into three stages. In the normal stage, the optical signal path loss increases slowly, for example, the monthly loss is ≤0.2 decibels. The loss rate in the early warning stage is significantly accelerated, for example, the monthly loss is 0.2-0.6 decibels. The loss rate in the fault stage rises sharply, for example, the monthly loss is ≥0.6 decibels. Each stage corresponds to a numerical interval, for example, the normal stage corresponds to the fusion feature matrix value 0-0.3, the early warning stage 0.3-0.7, and the fault stage 0.7-1.0. The values ​​in each interval are linearly converted. For example, the early warning stage (0.3-0.7) is mapped to the probability 0.4-0.8: if the fusion feature value of a node is 0.5, the initial failure probability is 0.6. Generate a matrix containing the initial failure probabilities of all nodes. For example, the fusion feature value of node A is 0.65 (fault stage) and the initial probability is 0.85.

[0201] 802. Dynamically adjust the distribution of the initial failure probability according to the deviation between the actual observation value and the initial probability, verify the rationality of the distribution, and perform physical constraint verification on the initial failure probability distribution;

[0202] In step 802, the physical constraint verification is to correct the initial probability based on the physical characteristics of the area where the node is located (such as material strength, environmental stress), and eliminate the prediction results that do not conform to the actual engineering.

[0203] In the embodiment of the present application, the rule engine and multi-source data fusion are used to realize probability correction, and the initial probability is compared with the actual historical failure record. For example, the initial probability of a node is 0.8, but there has been no failure in the past six months. The system determines that the initial probability is too high and is adjusted down to 0.64 according to the deviation ratio (such as 20%). If the compressive strength of the area where the node is located is lower than the safety threshold (such as concrete <30MPa), the probability is reduced by 30%. Environmental stress constraint: If the ambient humidity is >85% or the temperature is >50°C, the probability increases by 15% (due to the harsh environment that accelerates equipment aging). If the current load of the node exceeds 90% of the design value, the probability increases by 20%. The initial probability of node B is 0.7, but the compressive strength of the area where it is located is 25MPa (triggering rule 1), and the corrected probability is 0.7×0.7=0.49. The initial probability of node C is 0.6, and the ambient humidity reaches 90% (triggering rule 2), and the corrected probability is 0.6×1.15=0.69.

[0204] 803. Associate the failure probability after physical constraint verification with the dynamic parameters of the communication node, and generate a communication node failure probability with a unified benchmark based on the failure probability data between different communication nodes.

[0205] In step 803, dynamic parameter association associates the failure probability with the real-time operating parameters of the node (such as load rate, signal quality), ensuring that the probability value reflects the real-time status.

[0206] In the embodiment of the present application, dynamic weighting and normalization processing is used to achieve probability standardization. If the current load rate of the node is greater than 90%, the failure probability weight is increased by 20%. If the optical signal-to-noise ratio is less than 25dB, the weight is increased by 15%. If the temperature fluctuates by more than 5°C per hour, the weight is increased by 10%. The weighted probabilities of all nodes are normalized to the maximum and minimum, and the values ​​are compressed to the range of 0-1. For example, the original weighted probability range is 0.2-0.95, and the probability of node D changes from 0.8 to 0.84 after normalization. Generate a failure probability distribution map of nodes in the entire network, for example: node E probability 0.92 (red warning), node F probability 0.45 (yellow observation), node G probability 0.15 (green normal).

[0207] A smart grid optical communication system needs to predict the failure risk of key nodes in substations to prevent large-scale power outages. The fusion eigenvalue of node H is 0.68 (fault stage interval), which is mapped to an initial probability of 0.82 through a piecewise linear function. The area where node H is located is an outdoor substation with an ambient humidity of 88% (triggering rule 2), and the probability is increased to 0.82×1.15=0.94. However, it uses high-strength concrete (compressive strength 35MPa), which does not trigger rule 1, and the final probability remains 0.94. The current load rate of node H is 98% (weight + 20%), and the signal quality signal-to-noise ratio is 22dB (weight + 15%). The probability after weighted calculation is 0.94×1.35=1.27. The highest probability of the entire network is 1.27, the lowest is 0.1, and the final probability of node H is 1.0 (red emergency alarm). The system automatically isolates node H, switches to the backup line, and dispatches a maintenance work order. Actual inspection found that the optical cable connector was aging, and it was replaced in time to avoid power outages.

[0208] In summary, this solution improves the prediction accuracy of failure probability to more than 95% and reduces the false alarm rate to less than 5% through physical constraint verification and dynamic parameter association; in the smart grid scenario, it can achieve 72-hour early warning for high-risk nodes, improve operation and maintenance response efficiency by 50%, and shorten the fault repair time to within 2 hours. It is suitable for preventive maintenance of key infrastructure such as energy, transportation, and industry, and significantly reduces systemic risks and economic losses.

[0209] In order to improve the overall robustness and business continuity of the communication network, firstly, all candidate paths are traversed based on the optical cable topology database, and the three-dimensional Euclidean distance between each path and the high failure probability node is calculated through spatial geographic coordinates, and the candidate paths that meet the minimum geographic isolation threshold (such as ≥500 meters) are screened out; then, the path similarity detection algorithm (such as sequence alignment technology based on edit distance) is used to analyze the overlapping routing segments of the candidate paths and the main path, and the candidate paths that share pipelines, relay nodes or geographical areas with the main path exceeding a preset proportion (such as >20%) are eliminated; finally, a set of physically isolated and non-overlapping high-quality backup paths is generated by combining the path transmission performance parameters (such as the number of hops, delay difference) and engineering constraints (such as construction cost, maintenance difficulty), to ensure the independence and anti-destruction of the main and backup paths in extreme scenarios. In some embodiments, all candidate optical cable paths are traversed based on the time-space association relationship, and candidate optical cable paths that meet the minimum geographic isolation distance with the high failure probability node are screened out as backup optical cable paths, and candidate optical cable paths that have overlapping routes with the main optical cable path are excluded, including:

[0210] 901. Based on the time-space correlation relationship, perform transmission delay equalization processing on each candidate optical cable path in the candidate optical cable path set, eliminate the delay difference between the paths by dynamically adjusting the optical fiber dispersion compensation parameters, and generate a set of candidate optical cable path parameters after delay equalization;

[0211] In step 901, the transmission delay equalization process is to balance the transmission delay differences of different paths by dynamically adjusting the fiber dispersion compensation parameters (such as dispersion slope and compensation distance) of the optical cable path, so as to ensure the synchronization of the signal transmission time of the primary and backup paths. The candidate optical cable path parameter set includes a database of the physical parameters (such as length and material type) of each path, transmission performance (such as delay and jitter) and optimized dispersion compensation parameters.

[0212] In the embodiment of the present application, adaptive dispersion compensation technology is used to achieve delay equalization, and a high-precision optical time domain reflectometer (OTDR) is used to measure the transmission delay of each candidate path. For example, the delay of path A is 3.5 milliseconds, and that of path B is 4.2 milliseconds. The fiber dispersion parameters of the path (such as dispersion slope and nonlinear coefficient) are collected to determine the initial compensation value (such as the initial compensation of path A is 100ps / nm / km). High-latency path optimization For paths with higher delays (such as path B), the compensation amount of the dispersion compensation module is increased (for example, from 100ps / nm / km to 120ps / nm / km), and the delay is reduced to 3.8 milliseconds. Low-latency path fine-tuning For paths with lower delays (such as path A), the compensation amount is moderately reduced (for example, from 100ps / nm / km to 90ps / nm / km), so that the delay is slightly increased to 3.6 milliseconds, and the delay difference between the primary and backup paths is achieved ≤0.2 milliseconds. Generate a set of balanced path parameters. For example, the delay of path A is 3.6 ms, the delay of path B is 3.8 ms, and the delay difference of 0.2 ms meets the synchronization requirement.

[0213] 902. Performing geographic coordinate mapping on the candidate optical cable path parameter set after delay equalization, calculating the geographic isolation distance between adjacent optical cable paths, and selecting a backup optical cable path that meets the minimum geographic isolation distance with the high failure probability node;

[0214] In step 902, the geographic isolation distance is the minimum three-dimensional spatial distance between the backup path and the high failure probability node (e.g., horizontal distance ≥ 500 meters, vertical distance ≥ 10 meters), to prevent common disasters (earthquakes, floods) from simultaneously destroying the primary and backup paths. The optical cable path geographic coordinate mapping converts the physical route of the optical cable path (e.g., underground pipelines, relay station locations) into three-dimensional geographic coordinates (latitude and longitude, altitude), supporting spatial distance calculation.

[0215] In the embodiment of the present application, high-precision positioning and spatial distance calculation are used to achieve screening, and the Beidou positioning system is used to obtain the coordinates of the path nodes. For example, path A passes through the underground pipeline well P-01 (120.35° east longitude, 30.12° north latitude, -5 meters above sea level) to well P-05 (120.40° east longitude, 30.15° north latitude, -5 meters above sea level). Construct a three-dimensional spatial model of the path to support distance calculation and visual analysis. Calculate the Euclidean distance between the path node and the high-failure node. For example, node C (120.38° east longitude, 30.13° north latitude, -5 meters above sea level) is 550 meters away from the nearest pipeline well P-02 of path A, which meets the threshold of ≥500 meters. If a section of path D is only 300 meters away from node C, it is eliminated. Generate a list of candidate paths that meet the requirements of geographic isolation, for example, path A and path B are selected, and path D is excluded.

[0216] 903. Extract the optical cable links of the main path. If the backup optical cable path after noise suppression has the same optical cable link as the main path, it is determined to be a route overlapping path, and the candidate optical cable paths of the overlapping routes are excluded.

[0217] In step 903, the routing overlap path is the overlap between the backup path and the primary path in physical facilities (such as cable wells, relay stations) or geographical areas (such as the same street, pipeline corridor), which may cause a single point failure to affect the primary and backup paths.

[0218] In an embodiment of the present application, path sequence comparison and topology analysis are used to achieve exclusion and parse the physical links of the main path. For example, the main path uses cable wells J-01 to J-05 and relay station R-3. If the backup path uses the relay station R-3 or cable well J-03 of the main path, it is determined to be a facility-level overlap. If the backup path and the main path run parallel to each other in the underground corridor of the same street for more than 200 meters, it is determined to be a regional-level overlap. For example, backup path E uses cable well J-03 (accounting for 10%), triggering the facility overlap rule and being eliminated; backup path F uses independent cable wells J-06 to J-10 throughout, with no overlap, and is retained.

[0219] A cross-sea optical cable system needs to select backup paths during the typhoon season to ensure lossless service switching after the main path is interrupted. The main path has a delay of 4.0ms. Candidate paths G (delay 4.5ms) and H (delay 4.2ms) are equalized. Path G increases dispersion compensation to 125ps / nm / km, and the delay is reduced to 4.3ms; path H reduces compensation to 95ps / nm / km, and the delay is fine-tuned to 4.1ms. After equalization, the delay difference is 0.2ms (main path 4.0ms vs path H 4.1ms), which meets the synchronization requirements. Path G is 700 meters away from the high-risk node (undersea seismic zone), and path H is 550 meters away, both of which meet the geographical isolation threshold (≥500 meters). The main path uses submarine cable trenches T-01 to T-05, and path H partially passes through T-03 (accounting for 15%), triggering the overlap exclusion rule. Path G uses independent trenches T-06 to T-10 throughout the entire process, without overlap, and is finally selected. The typhoon caused the main path trench T-02 to collapse, and the system switched to path G within 50ms, with zero service interruption and a packet loss rate of less than 0.01%.

[0220] In summary, this solution improves the reliability of backup paths to 99.5% through delay balancing, geographical isolation screening, and route overlap elimination; in cross-sea optical cable scenarios, it achieves 50ms-level seamless switching, improves path independence and anti-destruction by 90%, and reduces the mis-cutting rate by less than 0.1%. It is suitable for high-risk scenarios such as submarine communications and base stations in earthquake zones, significantly reducing the risk of business interruption caused by natural disasters or equipment failures.

[0221] Figure 2 A structural diagram of a key information extraction system based on a multimodal model is provided for an embodiment of the present application, such as Figure 2 As shown, the system includes:

[0222] An acquisition module 21 acquires building structure stress distribution data and optical cable layout topology data, wherein the building structure stress distribution data includes dynamic deformation parameters of each area of ​​the building, and the optical cable layout topology data includes a temporal and spatial correlation between the physical position of the communication node and the optical signal transmission path;

[0223] A generation module 22 generates a failure probability of a communication node by fusing the dynamic deformation parameter with the historical attenuation characteristics of the optical signal transmission path in the spatiotemporal correlation relationship through a multimodal model;

[0224] The generation module 22 is also used to extract a backup optical cable path that is physically isolated from a high failure probability node from the time-space association relationship based on the failure probability, synchronously receive the primary and backup path optical signals through the space diversity reception technology, and perform cross-modal alignment on the time-frequency domain characteristics of the primary and backup path optical signals in combination with the multimodal model to generate a redundant signal set; divide the key information in the redundant signal set into data slices according to the node geographical distribution rule, associate the data slices with the optical signal wavelength characteristics of the remote node and the compressive strength parameters of the building structure through the multimodal model, and generate a multimodal hash identifier;

[0225] The reconstruction module 23, when detecting the failure of the target communication node, matches the data fragments having the same optical signal wavelength characteristics and located in the high compressive strength area from the remote node according to the multimodal hash identifier, and reconstructs the key information based on the complementary characteristics of the multipath signals in the redundant signal set.

[0226] Figure 2 The key information extraction system based on the multimodal model can be executed Figure 1 The implementation principle and technical effect of the key information extraction method based on a multimodal model described in the illustrated embodiment will not be described in detail. The specific manner in which each module and unit performs operations in the key information extraction system based on a multimodal model in the above embodiment has been described in detail in the embodiment of the method, and will not be described in detail here.

[0227] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A key information extraction method based on a multimodal model, characterized in that: include: Acquire building structure stress distribution data and optical cable layout topology data, wherein the building structure stress distribution data includes dynamic deformation parameters of each area of ​​the building, and the optical cable layout topology data includes the spatiotemporal correlation between the physical position of the communication node and the optical signal transmission path; Generate a failure probability of a communication node by fusing the dynamic deformation parameter with the historical attenuation characteristics of the optical signal transmission path in the spatiotemporal correlation relationship through a multimodal model; Based on the failure probability, a backup optical cable path physically isolated from the node with high failure probability is extracted from the time-space correlation relationship, the optical signals of the primary and backup paths are synchronously received through the space diversity receiving technology, and the time-frequency domain features of the primary and backup paths are cross-modally aligned in combination with the multimodal model to generate a redundant signal set; Divide the key information in the redundant signal set into data fragments according to the node geographical distribution rule, associate the data fragments with the optical signal wavelength characteristics of the remote nodes and the building structure compressive strength parameters through the multimodal model, and generate a multimodal hash identifier; When failure of the target communication node is detected, data fragments having the same optical signal wavelength characteristics and located in a high compressive strength area are matched from the remote node according to the multimodal hash identifier, and the key information is reconstructed based on the complementary characteristics of the multipath signals in the redundant signal set.

2. The method according to claim 1, characterized in that: According to the multimodal hash identifier, data fragments having the same optical signal wavelength characteristics and located in a high compressive strength area are matched from the remote node, and the key information is reconstructed based on the complementary characteristics of the multipath signals in the redundant signal set, including: Establishing a distributed hash mapping table in a remote node according to the multimodal hash identifier, performing a neighbor search on the hash mapping table, and screening and matching data fragments having the same optical signal wavelength characteristics based on the neighbor search; During the data shard matching process, the pressure sensor data of the remote nodes is collected in real time, and a dynamic matrix of node compressive strength is constructed. The compressive coefficient of each node is calculated through the dynamic matrix of node compressive strength, and nodes with large compressive coefficients are retained. Based on the nodes with large compressive coefficients, data shards located in the high compressive strength area are screened out; A space-frequency joint dictionary is established according to the complementary characteristics of the multipath signals in the redundant signal set, signal reconstruction error minimization and energy convergence conditions are output based on the space-frequency joint dictionary, and key information is output after convergence verification based on the signal reconstruction error minimization and energy convergence conditions.

3. The method according to claim 2, characterized in that The method of establishing a space-frequency joint dictionary according to the complementary characteristics of the multipath signals in the redundant signal set, outputting a signal reconstruction error minimization and energy convergence condition based on the space-frequency joint dictionary, and outputting key information after convergence verification based on the signal reconstruction error minimization and energy convergence condition, includes: Extract the spatiotemporal propagation characteristics of the complementary features of multipath signals from the redundant signal set, eliminate the phase ambiguity of the sensor array through space-frequency joint projection, and output the standardized multipath complementary feature tensor; The multi-path complementary feature tensor is optimized through orthogonal constraints to optimize the incoherence between basis functions and generate a space-frequency joint dictionary with path difference compensation characteristics. The reconstruction error objective function is constructed based on the space-frequency joint dictionary, and the second-order derivative matrix of the error function is optimized through the gradient descent strategy to output the signal reconstruction error minimization and energy convergence conditions; Based on the signal reconstruction error minimization and energy convergence conditions, a stable convergence path is screened through a double verification mechanism to output key information to resist multipath interference.

4. The method according to claim 1, characterized in that The failure probability of the communication node is generated by fusing the dynamic deformation parameter with the historical attenuation characteristics of the optical signal transmission path in the time-space correlation relationship through a multimodal model, including: The dynamic deformation parameter sequence and the spatiotemporal correlation are fused through a multimodal model, and the weight proportions of the two types of features in the optical signal transmission path are dynamically allocated to generate a fusion feature matrix. Performing multi-scale feature decomposition on the fused feature matrix to extract a multi-scale feature matrix containing steady-state attenuation components and transient jump components of different time granularities; Based on the multi-scale feature matrix, the feature contribution of each time granularity is adjusted by a dynamic weight optimization algorithm to generate an optimized fusion feature matrix; Performing state transition analysis on the optimized fusion feature matrix to calculate a piecewise linear function of the historical state attenuation feature of the optical signal transmission path; Based on the piecewise linear function, the numerical interval of the optimized fusion feature matrix is ​​mapped to the initial failure probability, the initial failure probability distribution is subjected to physical constraint verification, and the failure probability of the communication node is output.

5. The method according to claim 1, characterized in that Extracting a backup optical cable path physically isolated from the high failure probability node from the time-space correlation relationship, synchronously receiving the primary and backup path optical signals through the space diversity receiving technology, and performing cross-modal alignment on the time-frequency domain features of the primary and backup path optical signals in combination with the multimodal model to generate a redundant signal set, including: Based on the spatiotemporal association, all candidate optical cable paths are traversed to select candidate optical cable paths that meet the minimum geographical isolation distance with the high failure probability node as backup optical cable paths, and candidate optical cable paths that have overlapping routes with the main optical cable path are excluded; Diversity receiving devices are deployed at the receiving ends of the main optical cable path and the backup optical cable path to achieve synchronous reception of optical signals of the main and backup paths through spatial diversity receiving technology; The time domain waveform features and frequency domain spectrum features are extracted from the primary and backup path optical signals after noise suppression respectively and input into the multimodal model. The cross-modal alignment method is used to fuse the time and frequency domain features of the primary and backup path signals to generate a redundant signal set.

6. The method according to claim 1, characterized in that The multimodal model is used to associate the data fragment with the optical signal wavelength characteristics of the remote node and the compressive strength parameters of the building structure to generate a multimodal hash identifier, including: Based on the multimodal model, the optical signal of the remote node is subjected to frame and window processing, wherein the frame and window processing process is adaptively adjusted according to the building vibration frequency to construct an optical signal feature sequence; Mapping the compressive strength parameters of the building structure to a three-dimensional space grid coordinate system, calculating the dynamic compressive strength attenuation factor of each grid unit based on the three-dimensional space grid coordinate system, and generating a compressive strength characteristic tensor based on the compressive strength attenuation factor strength; A mapping relationship table between the optical signal feature sequence and the compressive strength feature tensor is established, and a joint feature matrix including wavelength and compressive strength coupling weights is generated based on the mapping relationship table; The correlation between the joint features is eliminated to generate a multimodal hash identifier with anti-collision characteristics.

7. The method according to claim 4, characterized in that Performing state transition analysis on the optimized fusion feature matrix to calculate the piecewise linear function of the historical state attenuation feature of the optical signal transmission path includes: Based on the position parameters in the fusion feature matrix, state features in the local neighborhood are extracted through a preset neighborhood set to generate a neighborhood state feature set; Performing time-varying weight assignment on the neighborhood state feature set, introducing an attenuation coefficient related to the time step for each neighborhood state, and dynamically adjusting the nonlinear attenuation gradient of the attenuation coefficient through a piecewise linear interpolation algorithm; State transition analysis is performed on the neighborhood state feature set after the attenuation coefficient is adjusted, the historical state is divided into multiple intervals, and the boundary value jump parameters across the intervals are calculated to generate a piecewise linear function of the attenuation characteristics of the historical state of the optical signal transmission path.

8. The method according to claim 4, characterized in that Based on the piecewise linear function, the numerical interval of the optimized fusion feature matrix is ​​mapped to the initial failure probability, the initial failure probability distribution is physically constrained and verified, and the failure probability of the communication node is output, including: Dividing the numerical interval of the fusion feature matrix into multiple sub-intervals based on the piecewise linear function, performing linear transformation on the numerical values ​​in the sub-intervals, and generating an initial failure probability; The initial failure probability is dynamically adjusted according to the deviation between the actual observation value and the initial probability, the rationality of the distribution is verified, and the physical constraint verification of the initial failure probability distribution is performed; The failure probability after physical constraint verification is associated with the dynamic parameters of the communication node, and based on the failure probability data between different communication nodes, a communication node failure probability with a unified benchmark is generated.

9. The method according to claim 5, characterized in that Based on the time-space association relationship, all candidate optical cable paths are traversed to select candidate optical cable paths that meet the minimum geographical isolation distance with the high failure probability node as the backup optical cable path, and candidate optical cable paths that have overlapping routes with the main optical cable path are excluded, including: Based on the time-space correlation relationship, each candidate optical cable path in the candidate optical cable path set is subjected to transmission delay equalization processing, and the delay difference between the paths is eliminated by dynamically adjusting the optical fiber dispersion compensation parameters, so as to generate a set of candidate optical cable path parameters after delay equalization; Performing geographic coordinate mapping on the candidate optical cable path parameter set after delay equalization, calculating the geographic isolation distance between adjacent optical cable paths, and screening out backup optical cable paths that meet the minimum geographic isolation distance with high failure probability nodes; The optical cable links of the main path are extracted. If there are optical cable links identical to the main path in the backup optical cable path after noise suppression, it is determined to be a route overlapping path, and the candidate optical cable paths of the overlapping routes are excluded.

10. A key information extraction system based on a multimodal model, characterized in that: include: An acquisition module is used to acquire building structure stress distribution data and optical cable layout topology data, wherein the building structure stress distribution data includes dynamic deformation parameters of each area of ​​the building, and the optical cable layout topology data includes the spatiotemporal correlation between the physical position of the communication node and the optical signal transmission path; A generation module, which generates a failure probability of a communication node by fusing the dynamic deformation parameter with the historical attenuation characteristics of the optical signal transmission path in the spatiotemporal correlation relationship through a multimodal model; A generation module, based on the failure probability, extracts a backup optical cable path physically isolated from the high failure probability node from the time-space association relationship, synchronously receives the primary and backup path optical signals through the space diversity reception technology, and performs cross-modal alignment on the time-frequency domain features of the primary and backup path optical signals in combination with the multimodal model to generate a redundant signal set; A generation module is provided to divide the key information in the redundant signal set into data fragments according to the node geographical distribution rule, and to associate the data fragments with the optical signal wavelength characteristics of the remote nodes and the compressive strength parameters of the building structure through the multimodal model to generate a multimodal hash identifier; The reconstruction module, when detecting the failure of the target communication node, matches the data fragments having the same optical signal wavelength characteristics and located in the high compressive strength area from the remote node according to the multimodal hash identifier, and reconstructs the key information based on the complementary characteristics of the multipath signals in the redundant signal set.

Images