Optical cable connection information management system based on ODF

Abstract

The invention relates to the technical field of optical communication engineering, and discloses an optical cable connection information management system based on an ODF. According to the system, multi-mode physical parameters are collected in real time and aligned in time and space through a multi-source sensing device integrated at a connection point, and a standardized connection data stream is generated. A continuing behavior characterization model and a loss characteristic evolution model are constructed based on the data stream, and a dynamic mapping relation between continuing tracks and quality indexes and a causal chain between continuing parameters and loss evolution are respectively disclosed. The system analyzes a deviation signal and a fluctuation mode in a continuing process in real time, generates a multi-target optimization strategy according to double models, and accurately maps a self-adaptive control instruction into a cooperative action sequence of continuing equipment, so that dynamic closed-loop adjustment of continuing operation parameters is realized. According to the system, full-dimension perception, intelligent diagnosis and precise adaptive control of the connection process can be realized, and the reliability and consistency of connection quality are improved.

Description

Technical Field

[0001] This invention relates to the field of optical communication engineering technology, specifically to an ODF-based optical cable splicing information management system. Background Technology

[0002] Current fiber optic splicing operations heavily rely on operator experience and the automation level of the splicing equipment. Mainstream splicing equipment typically features a limited number of sensors. The primary focus is on initial alignment accuracy and final loss estimation, with quality control logic based on a simple comparison of final measurement results with preset empirical thresholds. The massive, multi-dimensional, and dynamically changing physical parameters generated throughout the splicing process are not systematically collected and correlated for utilization.

[0003] Existing technical solutions have shortcomings. Due to the single sensor source and isolated data, the system cannot establish a dynamic correlation between the connection operation and the quality indicators in progress. This leads to a significant lag in quality assessment, allowing only passive judgment after connection is completed, and preventing intervention before defects occur. During the connection process, operators can only observe the input parameters and output results, lacking understanding of the physical mechanisms of the intermediate processes. When the connection effect is unsatisfactory, troubleshooting the root cause is extremely difficult, often requiring repeated trial and error.

[0004] Existing methods cannot capture and analyze transient anomaly signals during the splicing process in real time. Tiny deviation signals, due to their transient nature and low amplitude, are usually filtered out as noise; however, their cumulative effect or specific patterns can be a direct cause of increased splicing losses or decreased mechanical strength. Traditional control strategies, based on fixed rules, cannot identify these complex fluctuation patterns, let alone generate corresponding, fine-grained adaptive adjustment commands.

[0005] The industry needs a technical solution that enables transparent perception, intelligent diagnosis, and precise control of the entire splicing process. The purpose of this invention is to develop an optical cable splicing information management system capable of real-time insight into the underlying mechanisms of the splicing process and, based on this, proactively adjusting quality. Summary of the Invention

[0006] The purpose of this invention is to provide an ODF-based optical cable splicing information management system to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides an ODF-based optical cable splicing information management system, the system comprising:

[0008] The intelligent sensing module is used to capture the sequence of physical parameters during the optical cable splicing process in real time through a multi-source sensing device integrated at the splicing point, and to perform spatiotemporal alignment processing on the physical parameter sequence to generate a standardized splicing data stream.

[0009] The dynamic modeling module is used to construct a connection behavior representation model and a loss characteristic evolution model based on the standardized connection data stream. The connection behavior representation model describes the dynamic relationship between connection trajectory and quality indicators, and the loss characteristic evolution model describes the causal chain between connection parameters and loss changes.

[0010] The adaptive decision-making module is used to analyze the deviation signals and fluctuation patterns in the connection process in real time, generate a multi-objective optimization strategy based on the connection behavior representation model and the loss characteristic evolution model, and output adaptive control commands.

[0011] The precision execution module is used to map the adaptive control commands into the action sequence of the connecting device and dynamically adjust the operating parameters of the connecting device.

[0012] Preferably, the intelligent sensing module includes:

[0013] The multimodal sensing unit is used to simultaneously acquire deformation gradient data and temperature field distribution data at the splicing point through fiber optic strain sensors and thermal imagers deployed in the splicing area.

[0014] A streaming processing unit is used to perform sliding window analysis on the deformation gradient data and temperature field distribution data, and extract statistical feature vectors within the window.

[0015] The data normalization unit is used to map the statistical feature vector to a unified dimension space to eliminate the magnitude difference between sensors.

[0016] The feature enhancement unit uses a deep autoencoder to reduce the dimensionality and denoise the normalized features, generating an enhanced feature sequence.

[0017] The sequence generation unit is used to sort the enhanced feature sequences by timestamp and combine them into the standardized continuum data stream.

[0018] Preferably, the dynamic modeling module includes:

[0019] The trajectory learning component is used to input the continuous trajectory segments in the standardized continuous data stream into the gated recurrent unit network, learn the trajectory change pattern through the time series prediction algorithm, and generate the trajectory inference function in the continuous behavior representation model.

[0020] The loss learning component is used to input the loss fluctuation data in the standardized continuous data stream into the convolutional neural network, extract multi-scale loss patterns through the feature pyramid structure, and generate the loss propagation function in the loss characteristic evolution model.

[0021] A model fusion component is used to fuse the trajectory extrapolation function and the loss propagation function through a cross-attention mechanism to establish a joint representation space;

[0022] An online update component is used to adjust the weight matrices of the gated recurrent unit network and the convolutional neural network using incremental gradient descent based on new samples in the real-time continuous data stream.

[0023] The validation component is used to verify the model's generalization ability by calculating the reconstruction error between the joint representation space and real-time data.

[0024] Preferably, the adaptive decision-making module includes:

[0025] An anomaly detection unit is used to continuously monitor the offset of the splicing quality indicators and the oscillation amplitude of the loss sequence;

[0026] The trajectory decision unit is used to activate the trajectory inference function in the successive behavior representation model when the offset exceeds the dynamic threshold and the oscillation amplitude is in a stable range.

[0027] The trajectory instruction generation unit is used to calculate the correction vector of the successive trajectory based on the output of the trajectory deduction function, and generate trajectory optimization instructions.

[0028] A loss decision unit is used to activate the loss propagation function in the loss characteristic evolution model when the oscillation amplitude exceeds the tolerance range and the offset remains stable.

[0029] The loss command generation unit is used to determine the adjustment step size of the loss parameters based on the output of the loss propagation function and generate a loss control command.

[0030] The priority scheduling unit is used to prioritize the execution of trajectory optimization instructions and temporarily store loss control instructions until the trajectory stabilizes when both the offset and oscillation amplitude exceed the limits.

[0031] Preferably, the precision execution module includes:

[0032] The trajectory execution unit is used to parse the correction vector in the trajectory optimization instruction and decompose it into a motion increment sequence of each axis of the connecting device.

[0033] The trajectory monitoring unit is used to collect real-time deformation data of the connection point after each motion increment is executed, and to perform difference analysis with the predicted value of the trajectory extrapolation function.

[0034] The trajectory adjustment unit is used to maintain the current direction of motion until the target accuracy is reached if the difference value gradually converges; and to reverse the direction of motion and reinitialize the trajectory inference function if the difference value diverges.

[0035] The loss execution unit is used to parse the adjustment step size in the loss control command and modulate the output power curve of the connection device.

[0036] The loss monitoring unit is used to track splice loss changes in real time using an optical time-domain reflectometer and dynamically calibrate the parameters of the loss propagation function based on the tracking results.

[0037] Preferably, the system further includes a model calibration module, the model calibration module comprising:

[0038] The final state data acquisition unit is used to collect the final quality assessment data and loss profile data after the connection is completed;

[0039] The trajectory model comparison unit is used to calculate the matching degree between the final quality assessment data and the prediction interval of the subsequent behavior representation model, and generate a trajectory error distribution map.

[0040] The loss model registration unit is used to perform similarity analysis between the loss profile data and the expected template of the loss characteristic evolution model to generate a loss error spectrum.

[0041] The trajectory rule correction unit is used to extract systematic deviation components from the trajectory error distribution map and adjust the baseline parameters of the trajectory extrapolation function.

[0042] The loss parameter correction unit is used to identify random fluctuation components from the loss error spectrum and optimize the compensation coefficient of the loss propagation function.

[0043] Preferably, the trajectory rule correction unit includes:

[0044] The deviation decomposition component is used to separate the steady-state error and dynamic error in the trajectory error distribution map using the principal component analysis algorithm.

[0045] The trajectory reference adjustment component is used to correct the parameter settings of the trajectory extrapolation function based on the magnitude and direction of the steady-state error.

[0046] Preferably, the loss parameter correction unit includes:

[0047] The fluctuation filtering component is used to extract the effective fluctuation signal in the loss error spectrum using wavelet threshold denoising technology;

[0048] The loss weight adjustment component is used to reallocate the weights of each parameter in the loss propagation function according to the frequency characteristics of the effective fluctuation signal.

[0049] The model replacement component is used to replace the old version of the model with the updated trajectory extrapolation function and loss propagation function.

[0050] Preferably, the system further includes a preprocessing module, the preprocessing module comprising:

[0051] The optical cable parsing unit is used to decode the structural attribute code and material property segment in the model identifier of the target optical cable and generate the optical cable feature tensor.

[0052] The template retrieval unit is used to perform multi-dimensional similarity matching between the optical cable feature tensor and the preset optical cable knowledge base, and to filter out candidate splicing template groups with a matching degree higher than a threshold.

[0053] The performance calculation unit is used to extract the quality root mean square error and loss stability index from the historical connection records of each template in the candidate connection template group, and calculate the comprehensive performance score.

[0054] The template selection unit is used to sort the candidate template group in descending order according to the comprehensive performance score and select the optimal successor template;

[0055] The reference curve acquisition unit is used to retrieve the set of historical loss curves associated with the optimal splicing template from the optical cable knowledge base;

[0056] The collaborative analysis unit is used to perform trajectory-loss consistency checks on each curve in the historical loss curve set, remove abnormal curves with mismatch points, and form an optimized curve set.

[0057] The baseline curve determination unit is used to select the curve with the smallest distortion as the baseline loss curve based on the historical stability of each curve in the optimization curve set.

[0058] The parameter initialization unit is used to perform spatiotemporal registration of the optimal splicing template and the reference loss curve, and to set the initial splicing parameters.

[0059] Preferably, the collaborative analysis unit includes:

[0060] The data pairing component is used to select the curve to be tested from the historical loss curve set and simultaneously obtain the trajectory parameter sequence of the corresponding time point in the optimal continuation template;

[0061] The time alignment component is used to label the curve to be tested with collaborative time tags based on key event points in the trajectory parameter sequence;

[0062] The mutation identification component is used to scan the time-stamped loss curve, detect whether the loss gradient in the neighborhood of each label exceeds the critical value, and mark the mutation interval.

[0063] The collision detection component is used to compare the changing trend of trajectory parameters with the direction of the abrupt change in loss within the abrupt change range, and to calculate the collision intensity index.

[0064] The collision marker component is used to record the start and end times of a collision interval when the collision intensity index exceeds a threshold.

[0065] The curve reconstruction component is used to interpolate and generate smooth transition segments within the conflict interval based on historical conflict resolution cases.

[0066] The curve replacement component is used to cover the original conflict zone with a smooth transition segment, generating an optimized loss curve.

[0067] The integrity check component is used to verify the continuity of the optimized curve and filter out residual conflict points.

[0068] The set of building components is used to integrate all validated curves into an optimized curve set.

[0069] Compared with the prior art, the beneficial effects of the present invention are:

[0070] By integrating multi-source sensors at the connection point, heterogeneous physical parameter sequences, including optical images, thermal field distributions, electrical parameters, and mechanical data, are simultaneously acquired, and these data are processed using a spatiotemporal alignment algorithm. This algorithm unifies data streams with different sampling frequencies and physical dimensions to the same timestamp and spatial coordinate system, generating a standardized connection data stream with strict spatiotemporal correlation. This technique completely changes the traditional single-parameter recording mode, constructing a complete data image of the connection process, enabling precise correlation and retrospective analysis of the multidimensional physical state corresponding to each operation instant.

[0071] A continuity behavior representation model built upon standardized data streams establishes a dynamic mapping relationship between equipment operation trajectories and process quality indicators through machine learning methods. This model can learn characteristic patterns of excellent continuity operations from historical data and perform quality assessments on real-time continuity trajectories. Simultaneously, a loss characteristic evolution model employs causal inference techniques to reveal the causal transmission chain from abnormal process parameters to final loss formation. The collaborative work of these two models achieves a cognitive upgrade in continuity quality analysis, moving from correlation analysis to causal judgment.

[0072] During real-time continuity testing, the system continuously compares the actual acquired physical parameter sequences with the ideal trajectory predicted by the behavioral characterization model to identify minute deviation signals and specific fluctuation patterns. These transient features, ignored by traditional systems, become early indicators for predicting potential quality defects. Based on the deduction of deviation consequences using the loss characteristic evolution model, the system can calculate control strategies that simultaneously optimize multiple quality objectives, rather than performing simple compensation for a single objective.

[0073] The precision execution module transforms complex multi-objective optimization strategies into specific action sequences for the connecting equipment. This module employs a nonlinear mapping algorithm to parse strategy instructions into a precise instruction set for the coordinated operation of multiple actuators along the time axis. This achieves a shift from discrete parameter tuning to continuous process optimization. Attached Figure Description

[0074] Figure 1 This is a schematic diagram illustrating the working principle of the ODF-based optical cable splicing information management system described in this invention.

[0075] Figure 2 A flowchart illustrating how the intelligent sensing module works;

[0076] Figure 3 A flowchart illustrating the operation of the dynamic modeling module;

[0077] Figure 4 Analysis diagram of trajectory execution monitoring and adaptive adjustment process;

[0078] Figure 5 Figures show the loss error spectrum analysis and wavelet denoising processing. Detailed Implementation

[0079] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] Please see Figure 1 This invention provides an ODF-based optical cable splicing information management system, the overall implementation of which is as follows: The system integrates a multi-source sensing device at the optical cable splicing point with an intelligent sensing module to capture the sequence of physical parameters during the splicing process in real time. These parameters include, but are not limited to, physical quantities such as temperature, strain, and deformation. The intelligent sensing module performs spatiotemporal alignment processing on the original physical parameter sequence to eliminate the differences in acquisition time and space between different sensors, generating a standardized splicing data stream. This data stream is indexed by timestamps to ensure data consistency and processability. A dynamic modeling module receives the standardized splicing data stream and uses machine learning methods to construct a splicing behavior representation model and a loss characteristic evolution model. The splicing behavior representation model analyzes the dynamic correlation between the splicing trajectory and quality indicators to form a trajectory inference function, while the loss characteristic evolution model establishes a loss propagation function based on the causal relationship between splicing parameters and loss changes. The adaptive decision-making module monitors deviation signals and fluctuation patterns in real time during the splicing process. Deviation signals refer to the offset of quality indicators from expected values, and fluctuation patterns refer to the oscillating characteristics of the loss sequence. This module combines the outputs of the splicing behavior characterization model and the loss characteristic evolution model to generate a multi-objective optimization strategy. These objectives include minimizing splicing loss and maximizing splicing efficiency, and it generates adaptive control commands. The precision execution module translates the adaptive control commands into specific action sequences for the splicing equipment, such as adjusting the fiber alignment mechanism or modulating laser power, dynamically adjusting operating parameters to optimize the splicing process. The entire system operates in a closed-loop manner, achieving automated control from data acquisition to execution.

[0081] Example 1: See Figure 2 This embodiment details the specific implementation of the intelligent sensing module. This module captures the sequence of physical parameters during the optical cable splicing process in real time through multi-source sensing devices and performs spatiotemporal alignment processing on the physical parameter sequence to generate a standardized splicing data stream. In specific implementation, the multimodal sensing unit of the intelligent sensing module is responsible for deploying fiber optic strain sensors and thermal imagers in the splicing area. The fiber optic strain sensors are installed in a high-density array on the surface of the optical cable around the splicing point. Each sensor collects microscopic deformation data at a sampling rate of kilohertz, forming deformation gradient data. The deformation gradient data is stored in a tensor structure, containing timestamps and spatial coordinate information. The thermal imager uses infrared scanning technology to capture the temperature distribution of the splicing area at a rate of tens of frames per second, generating temperature field distribution data. The temperature field distribution data records the temperature value of each pixel in a grid format and is synchronized with the deformation gradient data through a hardware clock to ensure the synchronization of acquisition. The multimodal sensing unit transmits deformation gradient data and temperature field distribution data to the streaming processing unit in real time. The streaming processing unit is equipped with a circular buffer for temporary storage of incoming sensing data. The streaming processing unit applies a sliding window analysis algorithm, with the sliding window size set to millisecond-level time intervals, such as a 10-millisecond window. It performs real-time calculations on the deformation gradient data and temperature field distribution data within each window, extracting statistical feature vectors. These statistical feature vectors include the mean, standard deviation, peak value, and energy value of the data within the window. These features are output as vectors, representing the transient characteristics of the successive process. The data normalization unit receives the statistical feature vectors output by the streaming processing unit. The data normalization unit uses a min-max scaling method to map the statistical feature vectors generated by different sensors to a unified dimension space. For example, it normalizes the feature values ​​of deformation data to the range [0,1], and the feature values ​​of temperature data to the range [0,1], eliminating the magnitude influence caused by differences in sensor type. The normalization process is based on predefined scale parameters, which are learned from historical calibration data, ensuring the comparability of the normalized features. The feature enhancement unit employs a deep autoencoder network to reduce the dimensionality and denoise the normalized features. The deep autoencoder network consists of an encoder and a decoder. The encoder contains multiple fully connected layers that compress high-dimensional features into a low-dimensional latent space using a non-linear activation function, removing noise and redundant information. The decoder reconstructs the compressed features, ensuring key information is preserved, and outputs an enhanced feature sequence. This enhanced feature sequence is stored in a time-series format, offering higher signal-to-noise ratio and representativeness. The sequence generation unit sorts the enhanced feature sequences output by the feature enhancement unit by timestamps generated by the system's global clock, achieving microsecond-level accuracy. The sequence generation unit then combines the sorted enhanced feature sequences into a standardized contiguous data stream. This standardized contiguous data stream is encapsulated in a streaming data format, including a header and a payload. The header stores metadata such as the time range and sensor identifier, while the payload stores the feature sequence, facilitating real-time access and processing by subsequent modules.

[0082] In some embodiments, the fiber optic strain sensor of the multimodal sensing unit employs fiber Bragg grating technology. The fiber Bragg grating sensor is embedded in the sheath near the fiber optic cable splice point and arranged in a multi-point array. Each sensor measures strain changes and outputs a digital signal. The thermal imager uses an uncooled infrared detector, with its optical lens aligned with the splice area to generate a real-time thermal image sequence. This thermal image sequence is aligned with the strain data via timestamps. The sliding window analysis algorithm of the streaming processing unit employs an overlapping window strategy, with partial overlap between adjacent windows (e.g., 50%) to ensure the continuity of feature extraction. The calculation of statistical feature vectors uses an online algorithm to avoid storing large amounts of raw data and reduce memory usage. The normalization method of the data normalization unit is based on dynamic range adjustment. The dynamic range is updated according to the minimum and maximum values ​​of the real-time data stream, ensuring that normalization adapts to changes in the splice environment. The deep autoencoder network of the feature enhancement unit uses historical splice data for unsupervised learning during the training phase, with the training objective being to minimize reconstruction error. During the network inference phase, the deep autoencoder network directly processes real-time data and outputs an enhanced feature sequence. The timestamp sorting of the sequence generation units is based on a first-in-first-out queue, and the queue size is dynamically adjusted according to the data flow to prevent data loss; the format of the standardized continuation data stream adopts industry standards such as Apache Avro, which supports efficient serialization and deserialization.

[0083] Optionally, the multimodal sensing unit can integrate additional sensors such as humidity or vibration sensors to expand the range of physical parameter acquisition; however, the core implementation still relies primarily on fiber optic strain sensors and thermal imagers. The sliding window size of the streaming processing unit can be dynamically adjusted according to the characteristics of the splicing process; for example, a larger window can be used in the initial stage of splicing to capture macroscopic trends, while a smaller window can be used in the fine-tuning stage to capture detailed changes. Z-score normalization can be used as an alternative method for dimensionality mapping in the data normalization unit, but the min-max scaling method is preferred due to its simplicity and efficiency. The deep autoencoder network in the feature enhancement unit can be replaced with other dimensionality reduction techniques such as principal component analysis, but deep autoencoder networks are more suitable for complex splicing scenarios due to their nonlinear processing capabilities. A verification mechanism, such as cyclic redundancy checks, can be added to the data stream combination of the sequence generation unit to ensure data integrity.

[0084] It is understandable that the synchronous acquisition of the multimodal sensing unit relies on a high-precision clock synchronization protocol, implemented through a network time protocol or hardware triggering, to ensure the temporal consistency of deformation gradient data and temperature field distribution data. The statistical feature vector extraction of the streaming processing unit aims to capture the dynamic features of the successive processes, providing a foundation for subsequent modeling. The unified dimensionality space of the data normalization unit eliminates sensor differences, making data from different sources comparable. The dimensionality reduction and denoising process of the feature enhancement unit improves data quality and reduces the computational burden on subsequent modules. The timestamp sorting of the sequence generation unit ensures the correct temporal order of the data stream, avoiding processing errors caused by out-of-order processing.

[0085] In practical implementation, the arrangement of the fiber optic strain sensor array in the multimodal sensing unit is based on the optical cable structure design. The array spacing is determined according to the optical cable diameter and splicing accuracy requirements. For example, in a standard single-mode optical cable splice, the sensor spacing is set to the millimeter level. The resolution of the thermal imager is selected based on the size of the splicing area, typically using a 640x480 pixel resolution to balance detail and processing speed. The sliding window analysis algorithm of the streaming processing unit uses a streaming processing framework such as Apache Flink. The window triggering mechanism is based on event time or processing time to ensure real-time performance. The calculation of statistical feature vectors is parallelized and accelerated using multi-core processors. The normalization parameters of the data normalization unit are obtained through offline calibration. The calibration process uses a large amount of historical data to train the normalization model, and the model parameters are stored in a configuration file and loaded at runtime. The deep autoencoder network structure of the feature enhancement unit is designed as symmetrical, with the encoder and decoder having the same number of layers, and ReLU is used as the activation function. The training data covers various splicing conditions to enhance generalization ability. The sequence generation unit assembly process includes timestamp correction, and the correction algorithm is based on interpolation to handle clock drift and ensure the time accuracy of the data stream.

[0086] In some embodiments, the data acquisition frequency of the multimodal sensing unit can be adjusted according to the continuation phase, for example, using a high-frequency acquisition at the beginning of the continuation to capture rapid changes, and a low-frequency acquisition in the stabilization phase to save resources; however, the basic implementation maintains a fixed frequency to ensure consistency. The window size and overlap rate of the streaming processing unit can be dynamically set through a configuration file, allowing operators to adjust them according to actual needs. The dimensional mapping of the data normalization unit can be extended to multivariate normalization, handling multiple feature dimensions simultaneously, but univariate normalization is preferred due to its simplicity. The deep autoencoder network of the feature enhancement unit can introduce an attention mechanism to focus on important features, but the standard structure already meets most scenarios. The data stream format of the sequence generation unit can be compatible with multiple protocols, such as MQTT or Kafka, to adapt to different system integration requirements.

[0087] Optionally, the sensor deployment of the multimodal sensing unit can utilize wireless transmission to reduce wiring complexity, but wired connections are preferred due to their high reliability. The feature extraction algorithm of the streaming processing unit can incorporate frequency domain analysis such as Fast Fourier Transform, but time-domain statistical features are sufficient to describe sequential behavior. The normalization range of the data normalization unit can be customized, such as [0,1] or [-1,1], but the [0,1] range is commonly used due to its intuitiveness. Network training for the feature enhancement unit can be performed online to adapt to changes, but offline training ensures stability. The data stream of the sequence generation unit can be transmitted in chunks to handle large data volumes, but streaming transmission guarantees real-time performance.

[0088] Understandably, the overall design of the intelligent sensing module prioritizes real-time performance and accuracy. The multimodal sensing unit provides raw data, the streaming processing unit performs initial processing, the data normalization unit standardizes the scale, the feature enhancement unit optimizes data quality, and the sequence generation unit generates the final output. These steps are interconnected, forming a complete data processing chain. In practical implementation, the parameter settings for each unit are based on extensive experimental verification; for example, the sliding window size is determined through cross-validation, and the number of layers in the deep autoencoder network is optimized through grid search. However, the core logic remains unchanged: the conversion from multi-source sensors to a standardized data stream. Through this approach, the intelligent sensing module can reliably support subsequent processing in the optical cable splicing information management system.

[0089] Example 2: See Figure 3The dynamic modeling module constructs a continuity behavior representation model and a loss characteristic evolution model based on standardized continuity data streams. The adaptive decision-making module analyzes deviation signals and fluctuation patterns during the continuity process in real time to generate multi-objective optimization strategies. In specific implementation, the trajectory learning component of the dynamic modeling module inputs continuity trajectory segments from the standardized continuity data stream into a gated recurrent unit network. These continuity trajectory segments are continuous time-series data extracted from the data stream, representing the historical motion path of the continuity device. The gated recurrent unit network uses a time-series prediction algorithm to learn the trajectory change patterns. The time-series prediction algorithm uses historical trajectory data to train network weights, optimizes the loss function through backpropagation, and generates a trajectory prediction function. This trajectory prediction function can predict future trajectory points based on the current trajectory state, and its output is in the form of a mathematical function or a lookup table. The loss learning component inputs loss fluctuation data from a standardized continuation data stream into a convolutional neural network. This loss fluctuation data comes from real-time measurements by an optical temporal reflectometer and is stored in time-series format. The convolutional neural network is configured with a feature pyramid structure, which extracts loss patterns, including short-term fluctuations and long-term trends, through multi-scale convolutional layers. The filter sizes of the convolutional layers vary from fine-grained to coarse-grained to capture features at different time scales, generating a loss propagation function. This function describes the causal relationship between loss and continuation parameters, and the output is a parameterized model. The model fusion component fuses the trajectory inference function and the loss propagation function through a cross-attention mechanism. This mechanism calculates the correlation weights between the outputs of the trajectory inference function and the loss propagation function, based on dot-product attention, to establish a joint representation space. This joint representation space is a high-dimensional vector space that integrates trajectory and loss information to form a unified model for subsequent decision-making. The online update component adjusts the weight matrices of the gated recurrent unit network and the convolutional neural network using incremental gradient descent based on new samples in the real-time continuous data stream. Incremental gradient descent processes new data in mini-batch mode, and the learning rate is dynamically adjusted to prevent overfitting. The weight matrix update is based on the gradient direction. The validation component verifies the model's generalization ability by calculating the reconstruction error between the joint representation space and real-time data. The reconstruction error refers to the mean square error between the model's predicted value and the actual measured value. If the error exceeds a preset threshold, model retraining or adjustment is triggered.

[0090] In some embodiments, the gated recurrent unit network structure of the trajectory learning component includes multiple hidden layers, with the number of neurons in each hidden layer set according to the continuation complexity, and the activation function using the tanh function; the time series prediction algorithm uses a sliding window input, with the window size determined based on the continuation period. The feature pyramid structure of the convolutional neural network in the loss learning component includes downsampling and upsampling layers to process loss data at multiple resolutions; the output of the loss propagation function includes the loss prediction value and confidence interval. The cross-attention mechanism of the model fusion component uses a query-key-value model, where the query comes from the trajectory inference function, and the key and value come from the loss propagation function. The incremental gradient descent method combined with momentum optimization in the online update component accelerates the convergence process. The reconstruction error calculation of the validation component uses an online algorithm to monitor model performance in real time.

[0091] Optionally, the trajectory learning component can use a Long Short-Term Memory (LSTM) network instead of a Gated Recurrent Unit (GRU) network, but the GRU is preferred due to its fewer parameters. The feature pyramid structure of the loss learning component can be simplified to a single-scale convolution, but multi-scale processing better captures loss characteristics. The attention mechanism of the model fusion component can be replaced with serial or parallel fusion, but cross-attention provides more refined integration. The gradient descent method for the online update component can use stochastic gradient descent, but an incremental approach saves computational resources. The error threshold of the validation component can be dynamically adjusted based on historical error distributions.

[0092] It is understandable that the trajectory learning component's trajectory inference function learns the dynamic patterns of successive trajectories, providing a foundation for trajectory optimization; the loss learning component's loss propagation function models the loss change mechanism, supporting loss control. The joint representation space of the model fusion component enables collaborative analysis of trajectory and loss, improving the model's expressive power. The incremental learning of the online update component allows the model to adapt to changes in the successive process, maintaining accuracy. The reconstruction error monitoring of the validation component ensures model reliability.

[0093] In practical implementation, the anomaly detection unit of the adaptive decision-making module continuously monitors the offset of the connection quality index and the oscillation amplitude of the loss sequence. The offset of the connection quality index is calculated by comparing the real-time quality index with a preset benchmark value, and the offset is expressed in scalar form. The oscillation amplitude of the loss sequence is obtained by statistically analyzing the fluctuation range of the loss data, such as calculating the standard deviation or variance, and the oscillation amplitude is output in numerical form. When the offset exceeds the dynamic threshold and the oscillation amplitude is within a stable range, the trajectory decision-making unit activates the trajectory extrapolation function in the connection behavior representation model. The dynamic threshold is adaptively adjusted based on historical connection data, and the stable range is defined as the oscillation amplitude being below a certain threshold. The trajectory instruction generation unit calculates the correction vector of the connection trajectory based on the output of the trajectory extrapolation function. The correction vector represents the direction and magnitude of the trajectory adjustment and is generated through geometric transformation or optimization algorithms, outputting trajectory optimization instructions, which are encapsulated in a control command format. When the oscillation amplitude exceeds the tolerance range and the offset remains stable, the loss decision-making unit activates the loss propagation function in the loss characteristic evolution model. The tolerance range is preset by the system and determined based on the connection criteria. The loss instruction generation unit determines the adjustment step size of the loss parameters based on the output of the loss propagation function. The adjustment step size is calculated incrementally, for example, using the gradient descent step size, and generates loss control instructions, which include parameter adjustment values. The priority scheduling unit prioritizes the execution of trajectory optimization instructions when both the offset and oscillation amplitude exceed the limits, and temporarily stores the loss control instructions until the trajectory stabilizes. The priority is based on a trade-off between continuity safety and efficiency.

[0094] In some embodiments, the offset calculation of the anomaly detection unit uses a moving average filter to smooth the data and reduce the impact of noise; the oscillation amplitude analysis uses frequency domain transformation to identify periodic fluctuations. The dynamic threshold update of the trajectory decision unit uses an exponentially weighted moving average algorithm; the judgment of the stationary interval is based on statistical hypothesis testing. The correction vector calculation of the trajectory instruction generation unit involves a Jacobian matrix or an optimization solver. The tolerance range of the loss decision unit can be set in stages, with different tolerance values ​​for different successive stages. The adjustment step size of the loss instruction generation unit is determined through line search or a fixed ratio. The temporary storage mechanism of the priority scheduling unit uses a queue to manage the instruction execution order.

[0095] Optionally, the anomaly detection unit can incorporate machine learning anomaly detection algorithms such as Isolation Forest, but the threshold method is preferred due to its simplicity and real-time performance. The activation conditions of the trajectory decision unit can incorporate time constraints, but the basic logic remains unchanged. The vector computation of the trajectory instruction generation unit can be simplified to one-dimensional adjustment, but multi-dimensional vectors are more accurate. The tolerance range of the loss decision unit can be dynamically learned, but a preset value ensures stability. The step size of the loss instruction generation unit can adaptively change, but a fixed step size is easier to implement. The priority rules of the priority scheduling unit can be configured, but trajectory priority is the default strategy.

[0096] It is understandable that the real-time monitoring of the anomaly detection unit provides decision input, the trajectory decision unit and the loss decision unit select the model according to the conditions, the trajectory instruction generation unit and the loss instruction generation unit generate specific instructions, and the priority scheduling unit coordinates the instruction execution order to ensure that the system response is reasonable.

[0097] In practical implementation, the trajectory learning component of the dynamic modeling module matches the feature count of consecutive trajectory segments in the input layer dimension of the gated recurrent unit network, and generates a trajectory inference function in the output layer; the time series prediction algorithm is trained using historical data, and the training set covers various consecutive scenarios. The convolutional neural network of the loss learning component takes the loss time series as input, and the feature pyramid structure extracts features through convolution and pooling operations; the loss propagation function outputs the loss prediction model. The cross-attention mechanism of the model fusion component is implemented as a neural network layer, and the attention weights are learned. The incremental gradient descent method of the online update component updates the network parameters after each batch of new data. The reconstruction error calculation of the verification component uses online evaluation, and logs are recorded when the error exceeds the limit. The anomaly detection unit of the adaptive decision-making module integrates sensor data streams and calculates indicators in real time; the state machine management model of the trajectory decision-making unit and the loss decision-making unit is activated; the output format of the trajectory instruction generation unit and the loss instruction generation unit is standardized; and the priority scheduling unit uses a scheduling algorithm to manage the instruction queue.

[0098] In some embodiments, network training for the trajectory learning component can be accelerated using distributed computing, but single-machine training is sufficient for real-time systems; the feature pyramid of the loss learning component can have its layer count adjusted to balance accuracy and speed. The attention mechanism of the model fusion component can incorporate residual connections to improve training. The learning rate scheduling of the online update component uses a cosine annealing strategy. The error threshold of the validation component can be set in segments. The data sampling rate of the anomaly detection unit is adjustable, but real-time performance must be guaranteed. The threshold adjustment algorithm of the trajectory decision unit can be optimized, but the current method is effective. The vector computation of the trajectory instruction generation unit can cache the results to improve efficiency. The tolerance range of the loss decision unit can be externally configured. The step size of the loss instruction generation unit can be manually overridden. The queue of the priority scheduling unit can be implemented as a priority queue data structure.

[0099] Optionally, components of the dynamic modeling module can be deployed on edge devices to reduce latency, but cloud processing allows for more complex models; the decision logic of the adaptive decision-making module can be extended to a multi-agent system, but a centralized design simplifies implementation. Understandably, the entire implementation emphasizes real-time performance and reliability, with the dynamic modeling module providing accurate models, the adaptive decision-making module ensuring intelligent responses, and closed-loop control optimizing the succession process.

[0100] Example 3: The precision execution module is responsible for mapping adaptive control commands to the action sequence of the connecting device and dynamically adjusting the operating parameters. The model calibration module collects data after the connecting process for model optimization and calibration. In specific implementation, the trajectory execution unit of the precision execution module receives trajectory optimization commands from the adaptive decision module. The trajectory optimization commands include correction vector information. The trajectory execution unit parses the correction vector, which is a multi-dimensional mathematical vector representing the direction and magnitude of the connecting trajectory adjustment. The trajectory execution unit decomposes the correction vector into a motion increment sequence for each axis of the connecting device. The connecting device is usually a multi-axis precision robot. The motion increment sequence defines the displacement of each motion axis within a discrete time step, such as the increment values ​​of the X-axis, Y-axis, and Z-axis. The motion increment sequence is stored in array format and drives the motor to execute through pulse signals output by the motion control card. The trajectory monitoring unit immediately collects real-time deformation data of the connection point after each motion increment is executed. The real-time deformation data is provided in real time by fiber optic strain sensors, with a sampling frequency of up to kilohertz. The trajectory monitoring unit performs difference analysis between the real-time deformation data and the predicted value of the trajectory extrapolation function. The difference analysis calculates the difference between the actual deformation data and the predicted deformation data, and the difference value is used to evaluate the trajectory tracking accuracy. The trajectory adjustment unit makes decisions based on the changing trend of the difference value. If the difference value gradually converges (meaning the difference value monotonically decreases or approaches zero over time), the trajectory adjustment unit maintains the current motion direction until the target accuracy is reached, which is defined by the system's preset error tolerance. If the difference value diverges (meaning the difference value increases or oscillates more intensely over time), the trajectory adjustment unit reverses the motion direction and reinitializes the training process of the trajectory extrapolation function. Reversing the motion direction is achieved by inverting the motion increment. The reinitialization of the trajectory extrapolation function training process includes resetting the weights of the gated recurrent unit network and reloading historical training data. The loss execution unit parses the adjustment step size in the loss control command. The adjustment step size is a scalar value representing the amount of output power adjustment. The loss execution unit modulates the output power curve of the splicing device. The output power curve describes the power change over time in a functional form. The modulation process is achieved by changing the laser's drive current through a digital-to-analog converter. The loss monitoring unit tracks the splicing loss changes in real time using an optical time-domain reflectometer (OTDR). The OTD scans the optical cable with high-frequency pulses and measures the reflected signal to generate a loss curve. The loss monitoring unit dynamically calibrates the parameters of the loss propagation function based on the tracking results. The dynamic calibration uses a recursive least squares method to update the coefficients of the loss propagation function to reduce prediction errors.

[0101] In some embodiments, the motion increment sequence generation of the trajectory execution unit uses interpolation algorithms, such as linear interpolation or spline interpolation, to ensure smooth motion; the difference analysis of the trajectory monitoring unit can introduce filtering techniques, such as Kalman filtering, to reduce the impact of noise. The convergence judgment of the trajectory adjustment unit is based on the statistical difference values ​​within a sliding window, such as calculating the gradient of the difference values ​​within the window; the divergence judgment is made by monitoring the second derivative of the difference values. The power curve modulation of the loss execution unit can use a piecewise linear function to adapt to different succession stages; the parameter calibration of the loss monitoring unit can incorporate a forgetting factor to adapt to time-varying systems.

[0102] In practical implementation, the final-state data acquisition unit of the model calibration module is activated after the splicing operation is completed. It collects the final quality assessment data and loss profile data after splicing. The final quality assessment data includes the tensile strength test results and optical return loss measurements at the splicing point, stored in a structured format. The loss profile data comes from the final scan of the optical time-domain reflectometer, displaying the loss distribution curve near the splicing point. The trajectory model comparison unit calculates the matching degree between the final quality assessment data and the prediction interval of the splicing behavior characterization model. The prediction interval is defined by the output range of the trajectory derivation function at the confidence level. The matching degree calculation uses similarity metrics such as cosine similarity or Euclidean distance to generate a trajectory error distribution map, which displays the distribution of error values ​​along the splicing path in a two-dimensional chart. The loss model registration unit performs similarity analysis between the loss profile data and the expected template of the loss characteristic evolution model. The expected template is an ideal loss curve constructed based on historical successful splicing data. The similarity analysis uses a dynamic time warping algorithm or Pearson correlation coefficient to generate a loss error spectrum, which is a frequency domain representation that identifies the spectral characteristics of the loss deviation. The trajectory rule correction unit extracts systematic deviation components from the trajectory error distribution map. These systematic deviation components refer to recurring fixed patterns in the error, which are separated using principal component analysis or Fourier analysis. The trajectory rule correction unit adjusts the baseline parameters of the trajectory inference function, including the initial values ​​of the network bias and weight matrix. The loss parameter correction unit identifies random fluctuation components from the loss error spectrum. These random fluctuation components represent high-frequency noise or random variations in the error, extracted using wavelet transform or high-pass filtering. The loss parameter correction unit optimizes the compensation coefficients of the loss propagation function. These compensation coefficients are used to adjust the model's output offset and are updated using gradient descent.

[0103] In some embodiments, data collection by the final-state data acquisition unit can be automatically triggered, executing automatically when the connecting device enters an idle state; the matching degree calculation of the trajectory model comparison unit can be weighted, with higher weights given to errors in important intervals. Similarity analysis by the loss model registration unit can be performed at multiple scales to capture deviations at different accuracy levels; parameter adjustments by the trajectory rule correction unit can be constrained within a reasonable range to prevent overfitting. The optimization process of the loss parameter correction unit can verify generalization ability using cross-validation.

[0104] Optionally, the trajectory execution unit can integrate safety monitoring functions for emergency stopping when motion exceeds limits, but the basic implementation focuses on performance optimization; the difference analysis of the trajectory monitoring unit can be extended to multivariate analysis, but univariate difference values ​​are sufficient for control. The re-initialization process of the trajectory adjustment unit can be partially reset rather than fully reset to retain useful knowledge; the modulation of the loss execution unit can incorporate nonlinear compensation, but linear modulation simplifies implementation. The calibration of the loss monitoring unit can be performed periodically rather than in real time, but real-time calibration improves accuracy. The units of the model calibration module can run offline to reduce real-time load, but online calibration ensures timeliness.

[0105] Understandably, the precision execution module achieves precise adjustment of trajectory and loss through closed-loop control. The trajectory execution unit and loss execution unit translate instructions into actions, the trajectory monitoring unit and loss monitoring unit provide feedback, and the trajectory adjustment unit achieves adaptive optimization. The model calibration module improves model accuracy through post-processing analysis. The final-state data acquisition unit provides real data, the trajectory model comparison unit and loss model registration unit evaluate model performance, and the trajectory rule correction unit and loss parameter correction unit complete model iteration.

[0106] In practical implementation, the difference analysis of the trajectory monitoring unit uses a mathematical formula to calculate the difference value, which is expressed as:

[0107]

[0108] in: The value representing the difference is a non-negative real scalar. The number of dimensions representing the deformation data is a positive integer. The actual deformation measurement value representing the i-th dimension is a real number; The predicted value of the trajectory extrapolation function representing the i-th dimension is a real number. This formula calculates the Euclidean distance between the actual deformation vector and the predicted deformation vector, serving as a quantitative indicator for difference analysis. The trajectory adjustment unit makes decisions based on the changing trend of the difference value D. If D monotonically decreases within a continuous time step, it is considered convergent; if D increases or fluctuates beyond a threshold within a continuous time step, it is considered divergent. When adjusting the baseline parameters of the trajectory extrapolation function, the trajectory rule correction unit refers to the systematic deviation in the trajectory error distribution map. The extraction of the systematic deviation component is achieved by fitting the trend line of the error distribution. When optimizing the compensation coefficient of the loss propagation function, the loss parameter correction unit uses the statistical characteristics of the random fluctuation component, such as variance or entropy, to adjust the coefficient magnitude.

[0109] In some embodiments, the formula for the trajectory monitoring unit can be replaced with other norm calculations, such as Manhattan distance, but Euclidean distance is preferred due to its universality; the decision logic of the trajectory adjustment unit can add a lag interval to prevent frequent switching. Data processing in the model calibration module can be parallelized, utilizing multi-core processors to accelerate analysis.

[0110] It is understandable that the collaborative work of the precision execution module and the model calibration module forms a complete control loop. The precision execution module ensures real-time optimization of the process, while the model calibration module guarantees long-term model reliability, thereby improving the overall system performance through a data-driven approach. In practical implementation, the interface design of all units follows a modular principle, allowing for independent testing and upgrades; parameter configuration is managed through configuration files, supporting flexible adjustments.

[0111] See Figure 4 In the graph, the solid blue line represents the actual deformation measurement value collected in real time by the fiber optic strain sensor, reflecting the true deformation state of the connection point during the precision robot's movement. The dashed red line shows the predicted deformation value generated by the trajectory extrapolation function based on the instructions of the adaptive decision module, demonstrating the model's mathematical extrapolation capability for the connection trajectory. The green curve shows the change in the difference between the actual deformation and the predicted deformation. This difference value is calculated using the Euclidean distance formula and is a key indicator for the trajectory adjustment unit's decision-making. When the difference value shows a monotonically decreasing trend, the system determines it to be in a convergent state, and continues to execute the current motion direction. When the difference value increases or fluctuates drastically, the system determines it to be in a divergent state, triggering a motion direction reversal and model re-initialization process. The triangles and inverted triangles marked in the graph represent the convergence and divergence decision points identified by the system, respectively. These decision points demonstrate the trajectory adjustment unit's intelligent judgment capability based on the changing trend of the difference value. The entire monitoring process is performed at a kilohertz sampling frequency, ensuring high accuracy and real-time performance of trajectory tracking.

[0112] Example 4: This example describes the specific implementation of the trajectory rule correction unit and the loss parameter correction unit in the model calibration module. The trajectory rule correction unit extracts systematic deviation components from the trajectory error distribution map and adjusts the baseline parameters of the trajectory extrapolation function. The loss parameter correction unit identifies random fluctuation components from the loss error spectrum and optimizes the compensation coefficients of the loss propagation function. Finally, the model update is completed through the model replacement component. In specific implementation, the deviation decomposition component of the trajectory rule correction unit uses principal component analysis (PCA) to separate the steady-state error and dynamic error in the trajectory error distribution map. The trajectory error distribution map is a two-dimensional data matrix generated by the trajectory model comparison unit, containing the error values ​​of each point on the connecting path. PCA calculates the covariance matrix of the error data, extracts eigenvectors and eigenvalues, and decomposes the error into principal and minor components. The steady-state error corresponds to the principal component with a larger eigenvalue, representing a long-term systematic deviation, while the dynamic error corresponds to the principal component with a smaller eigenvalue, representing short-term fluctuations or noise. The output of the deviation decomposition component is the separated error components, stored in vector form for subsequent adjustment. The trajectory reference adjustment component corrects the parameter settings of the trajectory extrapolation function based on the magnitude and direction of the steady-state error. The magnitude and direction of the steady-state error are calculated using the magnitude and direction angle of the error vector. The trajectory reference adjustment component accesses the parameter library of the trajectory extrapolation function, which stores the weights and biases of the gated recurrent unit network. The adjustment process uses gradient descent to fine-tune the parameters to reduce the impact of the steady-state error. After completing the adjustment, the trajectory reference adjustment component generates an updated version of the trajectory extrapolation function.

[0113] In some embodiments, the principal component analysis algorithm of the deviation decomposition component can be configured to retain the number of principal components, for example, retaining the top k principal components to capture most of the variance; the parameter adjustment of the trajectory baseline adjustment component can be constrained within a preset range to prevent over-adjustment. Optionally, the deviation decomposition component can use independent component analysis instead of principal component analysis to handle non-Gaussian errors, but principal component analysis is preferred due to its high computational efficiency; the trajectory baseline adjustment component can introduce a regularization term to avoid overfitting, but the basic adjustment method is effective enough.

[0114] In practical implementation, the fluctuation filtering component of the loss parameter correction unit extracts the effective fluctuation signal from the loss error spectrum using wavelet threshold denoising technology. The loss error spectrum is frequency domain data generated by the loss model registration unit, representing the spectral distribution of the loss deviation. Wavelet threshold denoising technology selects appropriate wavelet basis functions, such as the Daubechies wavelet, to perform multi-resolution analysis on the loss error spectrum. Wavelet coefficients are processed using soft or hard thresholding to remove high-frequency noise and retain the effective fluctuation signal. The output of the fluctuation filtering component is the denoised fluctuation signal, stored in time-series format. The loss weight adjustment component reallocates the weights of each parameter in the loss propagation function based on the frequency characteristics of the effective fluctuation signal. These frequency characteristics are obtained through Fourier transform or power spectrum analysis to identify the dominant frequency components. The loss weight adjustment component accesses the parameter table of the loss propagation function, which stores the weight values ​​of each layer in the convolutional neural network. It reallocates the contribution based on the frequency components; for example, high-frequency fluctuations correspond to rapidly changing parameters, and low-frequency fluctuations correspond to slowly changing parameters. The weight allocation is adjusted using a weighted average method or optimization algorithm. After completing the adjustment, the loss weight adjustment component generates an updated version of the loss propagation function. The model replacement component replaces the old model with the updated trajectory extrapolation function and loss propagation function. It checks the compatibility and performance of the new model and manages model files through a version control mechanism to ensure a seamless switch to the new model. See Table 1 for the correspondence between frequency characteristics and parameter weight adjustments in the loss error spectrum.

[0115] Table 1: Mapping Table of Frequency Characteristics and Parameter Weights

[0116] Frequency range (Hz) Fluctuation type description Corresponding parameter category Weighting adjustment factor 0-0.1 Low-frequency slow fluctuations Long-term trend parameters 1.2 0.1-1 Mid-frequency periodic fluctuations Periodic adjustment parameters 0.8 1-10 High-frequency random fluctuations Noise suppression parameters 0.5 10 or more Ultra-high frequency transient fluctuations Instantaneous response parameters 0.3

[0117] In practical implementation, the wavelet thresholding denoising technique of the fluctuation filtering component requires setting threshold parameters, which are adaptively determined based on the statistical characteristics of historical error data. The weight redistribution process of the loss weight adjustment component can be iteratively performed until the fluctuation signal stabilizes. The model replacement component uses timestamps or hash values ​​to identify model versions, and backups are performed before replacement to prevent data loss. Optionally, the fluctuation filtering component can be combined with Kalman filtering for preprocessing, but wavelet thresholding denoising is preferred due to its strong ability to handle non-stationary signals. The loss weight adjustment component can introduce machine learning methods to automatically learn weight mappings, but frequency-based rule-based methods are more transparent. The model replacement component can support hot replacement, i.e., updating the model without interrupting system operation, but cold replacement is safer.

[0118] It is understandable that the trajectory rule correction unit reduces system errors and improves trajectory prediction accuracy through deviation decomposition and benchmark adjustment; the loss parameter correction unit enhances model robustness through fluctuation filtering and weight optimization; and the model replacement component ensures the reliability of the model update process. The entire implementation emphasizes meticulous data processing and dynamic parameter adjustment to improve the model's adaptability during splicing. In specific implementation, the parameter settings of all components are calibrated based on a large amount of experimental data; for example, the selection of wavelet bases is determined through cross-validation, and the calculation of weight adjustment factors is based on the principle of error minimization. However, the core logic remains unchanged, namely, closed-loop optimization from error analysis to model correction. Through the above methods, this embodiment can effectively support the long-term performance maintenance of the optical cable splicing information management system.

[0119] See Figure 5 The blue curve in the figure shows the original loss error spectrum, containing complete fluctuation information from low to high frequencies. This error spectrum is generated by the loss model registration unit and reflects the frequency domain representation of the deviation between the actual loss profile data and the expected template. The red curve shows the effective fluctuation signal after wavelet threshold denoising. This process uses Daubechies wavelet basis functions for multi-resolution analysis and effectively separates noise components and useful information in the signal through soft thresholding. The different colored filled areas in the graph identify four key frequency ranges: the low-frequency slow fluctuation region mainly corresponds to the changes in long-term trend parameters; the mid-frequency periodic fluctuation region reflects the fluctuation characteristics of periodically adjusted parameters; the high-frequency random fluctuation region reflects the influence of noise suppression parameters; and the ultra-high-frequency transient fluctuation region represents the rapid changes in instantaneous response parameters. This identification of frequency characteristics provides an important basis for the parameter weight reallocation of the loss weight adjustment unit. By extracting the effective fluctuation signal through wavelet denoising, the system can more accurately identify the frequency characteristics of each parameter in the loss propagation function and then reallocate the parameter weights based on the contribution of frequency components. High-frequency fluctuations correspond to rapidly changing parameters, while low-frequency fluctuations correspond to slowly changing parameters. This refined weight allocation strategy significantly improves the accuracy and robustness of the loss characteristic evolution model.

[0120] Example 5: The preprocessing module parses the target optical cable and initializes the splicing parameters before splicing begins. In specific implementation, the optical cable parsing unit of the preprocessing module decodes the structural attribute code and material characteristic segment in the model identifier of the target optical cable. The model identifier is a unique identification code provided by the optical cable manufacturer, usually existing in the form of a string or QR code on the outer sheath of the optical cable. The structural attribute code describes the internal structural layers of the optical cable, such as the number of fiber cores, the type of reinforcement, and the number of sheath layers. The material characteristic segment indicates the material composition of each component of the optical cable, such as the optical fiber material, the type of filling compound, and the armor material. The optical cable parsing unit extracts this encoded information through a parsing algorithm, which includes string segmentation, regular expression matching, and database querying. The decoding results are combined into an optical cable feature tensor, which is a multi-dimensional numerical array, with each dimension representing an optical cable characteristic parameter, such as the numerical representation of the fiber core diameter, the material stiffness coefficient, and the structural complexity index. The template retrieval unit performs multi-dimensional similarity matching between the optical cable feature tensor and a pre-defined optical cable knowledge base. This knowledge base is a relational database storing a large amount of historical optical cable model splicing template data. Multi-dimensional similarity matching calculates the similarity score between the optical cable feature tensor and the feature vector of each template in the knowledge base. The similarity score is calculated using a cosine similarity algorithm or the reciprocal of Euclidean distance. Candidate splicing template groups with a matching degree higher than a pre-defined threshold are selected. Each candidate splicing template group contains multiple splicing template records, and each record contains complete splicing parameter settings. The performance calculation unit extracts the quality mean squared error and loss stability index from the historical splicing records of each template in the candidate splicing template group. The quality mean squared error measures the consistency of the template's quality indicators in historical splicing, while the loss stability index reflects the fluctuation of loss values ​​after splicing. The performance calculation unit calculates a comprehensive performance score based on the quality mean squared error and loss stability index. The comprehensive performance score is obtained through a weighted summation formula, with the weighting coefficients dynamically adjusted based on splicing requirements. The template selection unit sorts the candidate template groups in descending order based on the comprehensive performance score and selects the optimal splicing template, which is the template record with the highest comprehensive performance score. The reference curve acquisition unit retrieves a set of historical loss curves associated with the optimal splicing template from the optical cable knowledge base. This set contains multiple time-series data, each recording the loss changes during a historical splicing process. The collaborative analysis unit performs a trajectory-loss consistency check on each curve in the historical loss curve set, comparing the synchronization between the loss curve and the corresponding splicing trajectory parameters. The collaborative analysis unit removes abnormal curves with mismatch points, forming an optimized curve set containing loss curve data that passed the check. The baseline curve determination unit selects the curve with the lowest distortion as the baseline loss curve based on the historical stability of each curve in the optimized curve set. Historical stability is evaluated by calculating the curve's smoothness and continuity indicators, while distortion is quantified using a signal processing algorithm.The parameter initialization unit performs spatiotemporal registration of the optimal splicing template and the reference loss curve. The spatiotemporal registration aligns the time and space coordinates and sets the initial splicing parameters, which include the operating parameters such as the movement speed of the splicing equipment, the docking pressure, and the energy output.

[0121] In some embodiments, the model identifier decoding of the optical cable parsing unit can support multiple encoding formats, such as barcodes, RFID tags, or text descriptions, but the core parsing logic remains consistent. The similarity matching threshold of the template retrieval unit can be dynamically adjusted according to the splicing environment to improve matching accuracy. The comprehensive performance score calculation of the performance calculation unit can incorporate more indicators, such as splicing time efficiency, but quality and loss indicators are sufficient for basic implementation. Optionally, the optical cable parsing unit can integrate optical character recognition technology to process printed identifiers, but directly decoding digital identifiers is more reliable. The template retrieval unit can use an approximate nearest neighbor algorithm to accelerate large-scale knowledge base searches, but exact matching ensures accuracy. The weight coefficients of the performance calculation unit can be dynamically optimized based on machine learning models, but fixed weights simplify implementation.

[0122] In practical implementation, the data pairing component of the collaborative analysis unit selects the curve to be tested from the historical loss curve set and simultaneously acquires the trajectory parameter sequence at the corresponding time point in the optimal splicing template. The trajectory parameter sequence includes the position, velocity, and acceleration data of the splicing device. The time alignment component marks collaborative time tags on the curve to be tested based on the key event points of the trajectory parameter sequence. Key event points refer to significant changes in the trajectory, such as the splice start point, fiber contact point, and fusion splice point. The time alignment component uses an interpolation algorithm to ensure that the trajectory parameters and loss data are synchronized in time. The mutation identification component scans the time-tagged loss curves and detects whether the loss gradient in the neighborhood of each tag exceeds a critical value, which is set based on historical data statistics. The mutation identification component marks mutation intervals, which are the time periods when the loss gradient exceeds the critical value. The conflict detection component compares the change trend of the trajectory parameters with the direction of the loss mutation within the mutation interval and calculates the conflict intensity index, which is obtained by comparing the correlation between the trajectory change rate and the loss change rate. The conflict marking component records the start and end times of the conflict interval when the conflict intensity index exceeds a threshold, which is preset according to the splicing quality standard. The curve reconstruction component generates a smooth transition segment within the conflict interval based on historical conflict resolution cases, using either spline interpolation or linear interpolation algorithms. The curve replacement component covers the original conflict interval with the smooth transition segment to generate an optimized loss curve. The integrity check component verifies the continuity of the optimized curve and filters out residual conflict points; the continuity check is achieved by calculating the first derivative continuity of the curve. The set construction component integrates all verified curves into an optimized curve set.

[0123] In some embodiments, the data pairing component can process multiple curves in parallel to improve efficiency; the time synchronization accuracy of the time alignment component can reach the millisecond level. The gradient calculation of the mutation identification component can use the sliding window difference method; the correlation calculation of the conflict detection component can use the Pearson correlation coefficient. Optionally, the curve reconstruction component can select different interpolation methods to adapt to the curve characteristics; the integrity check component can set multi-level verification criteria. The set construction component can support incremental updates to optimize the curve set.

[0124] Understandably, the preprocessing module lays the foundation for the splicing process through system data preparation and parameter initialization. The optical cable analysis unit identifies optical cable characteristics, the template retrieval unit finds matching templates, the performance calculation unit evaluates template performance, the template selection unit determines the optimal selection, the reference curve acquisition unit provides historical data, the collaborative analysis unit optimizes data quality, the baseline curve determination unit selects reference standards, and the parameter initialization unit completes the final settings. The sub-components of the collaborative analysis unit ensure the consistency between the loss curve and trajectory data, providing reliable input for subsequent splicing. The entire implementation emphasizes data accuracy and consistency, ensuring optimal initial splicing state through multi-step processing. In specific implementations, all component parameters and algorithms undergo rigorous testing; for example, the similarity threshold is determined through receiver operation characteristic curve analysis, and the conflict detection threshold is set based on statistical significance, ensuring the system's reliability in practical applications.

[0125] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for managing optical fiber splicing information based on ODF, characterized in that, The system includes: The intelligent sensing module is used to capture the sequence of physical parameters during the optical cable splicing process in real time through a multi-source sensing device integrated at the splicing point, and to perform spatiotemporal alignment processing on the physical parameter sequence to generate a standardized splicing data stream. The dynamic modeling module is used to construct a connection behavior representation model and a loss characteristic evolution model based on the standardized connection data stream. The connection behavior representation model describes the dynamic relationship between connection trajectory and quality indicators, and the loss characteristic evolution model describes the causal chain between connection parameters and loss changes. The adaptive decision-making module is used to analyze the deviation signals and fluctuation patterns in the connection process in real time, generate a multi-objective optimization strategy based on the connection behavior representation model and the loss characteristic evolution model, and output adaptive control commands. The precision execution module is used to map the adaptive control commands into the action sequence of the connecting device and dynamically adjust the operating parameters of the connecting device.

2. The ODF-based optical cable splicing information management system according to claim 1, characterized in that, The intelligent sensing module includes: The multimodal sensing unit is used to simultaneously acquire deformation gradient data and temperature field distribution data at the splicing point through fiber optic strain sensors and thermal imagers deployed in the splicing area. A streaming processing unit is used to perform sliding window analysis on the deformation gradient data and temperature field distribution data, and extract statistical feature vectors within the window. The data normalization unit is used to map the statistical feature vector to a unified dimension space to eliminate the magnitude difference between sensors. The feature enhancement unit uses a deep autoencoder to reduce the dimensionality and denoise the normalized features, generating an enhanced feature sequence. The sequence generation unit is used to sort the enhanced feature sequences by timestamp and combine them into the standardized continuum data stream.

3. The ODF-based optical cable splicing information management system according to claim 2, characterized in that, The dynamic modeling module includes: The trajectory learning component is used to input the continuous trajectory segments in the standardized continuous data stream into the gated recurrent unit network, learn the trajectory change pattern through the time series prediction algorithm, and generate the trajectory inference function in the continuous behavior representation model. The loss learning component is used to input the loss fluctuation data in the standardized continuous data stream into the convolutional neural network, extract multi-scale loss patterns through the feature pyramid structure, and generate the loss propagation function in the loss characteristic evolution model. A model fusion component is used to fuse the trajectory extrapolation function and the loss propagation function through a cross-attention mechanism to establish a joint representation space; An online update component is used to adjust the weight matrices of the gated recurrent unit network and the convolutional neural network using incremental gradient descent based on new samples in the real-time continuous data stream. The validation component is used to verify the model's generalization ability by calculating the reconstruction error between the joint representation space and real-time data.

4. The ODF-based optical cable splicing information management system according to claim 3, characterized in that, The adaptive decision-making module includes: An anomaly detection unit is used to continuously monitor the offset of the splicing quality indicators and the oscillation amplitude of the loss sequence; The trajectory decision unit is used to activate the trajectory inference function in the successive behavior representation model when the offset exceeds the dynamic threshold and the oscillation amplitude is in a stable range. The trajectory instruction generation unit is used to calculate the correction vector of the successive trajectory based on the output of the trajectory deduction function, and generate trajectory optimization instructions. A loss decision unit is used to activate the loss propagation function in the loss characteristic evolution model when the oscillation amplitude exceeds the tolerance range and the offset remains stable. The loss command generation unit is used to determine the adjustment step size of the loss parameters based on the output of the loss propagation function and generate a loss control command. The priority scheduling unit is used to prioritize the execution of trajectory optimization instructions and temporarily store loss control instructions until the trajectory stabilizes when both the offset and oscillation amplitude exceed the limits.

5. The ODF-based optical cable splicing information management system according to claim 4, characterized in that, The precise execution module includes: The trajectory execution unit is used to parse the correction vector in the trajectory optimization instruction and decompose it into a motion increment sequence of each axis of the connecting device. The trajectory monitoring unit is used to collect real-time deformation data of the connection point after each motion increment is executed, and to perform difference analysis with the predicted value of the trajectory extrapolation function. The trajectory adjustment unit is used to maintain the current direction of motion until the target accuracy is reached if the difference value gradually converges; and to reverse the direction of motion and reinitialize the trajectory inference function if the difference value diverges. The loss execution unit is used to analyze the adjustment step size in the loss control command and modulate the output power curve of the connection device. The loss monitoring unit is used to track splice loss changes in real time using an optical time-domain reflectometer and dynamically calibrate the parameters of the loss propagation function based on the tracking results.

6. The ODF-based optical cable splicing information management system according to claim 1, characterized in that, The system also includes a model calibration module, which includes: The final state data acquisition unit is used to collect the final quality assessment data and loss profile data after the connection is completed; The trajectory model comparison unit is used to calculate the matching degree between the final quality assessment data and the prediction interval of the subsequent behavior representation model, and generate a trajectory error distribution map. The loss model registration unit is used to perform similarity analysis between the loss profile data and the expected template of the loss characteristic evolution model to generate a loss error spectrum. The trajectory rule correction unit is used to extract systematic deviation components from the trajectory error distribution map and adjust the baseline parameters of the trajectory extrapolation function. The loss parameter correction unit is used to identify random fluctuation components from the loss error spectrum and optimize the compensation coefficient of the loss propagation function.

7. The ODF-based optical cable splicing information management system according to claim 6, characterized in that, The trajectory rule correction unit includes: The deviation decomposition component is used to separate the steady-state error and dynamic error in the trajectory error distribution map using the principal component analysis algorithm. The trajectory reference adjustment component is used to correct the parameter settings of the trajectory extrapolation function based on the magnitude and direction of the steady-state error.

8. The ODF-based optical cable splicing information management system according to claim 7, characterized in that, The loss parameter correction unit includes: The fluctuation filtering component is used to extract the effective fluctuation signal in the loss error spectrum using wavelet threshold denoising technology; The loss weight adjustment component is used to reallocate the weights of each parameter in the loss propagation function according to the frequency characteristics of the effective fluctuation signal. The model replacement component is used to replace the old version of the model with the updated trajectory extrapolation function and loss propagation function.

9. The ODF-based optical cable splicing information management system according to claim 1, characterized in that, The system further includes a preprocessing module, which includes: The optical cable parsing unit is used to decode the structural attribute code and material property segment in the model identifier of the target optical cable and generate the optical cable feature tensor. The template retrieval unit is used to perform multi-dimensional similarity matching between the optical cable feature tensor and the preset optical cable knowledge base, and to filter out candidate splicing template groups with a matching degree higher than a threshold. The performance calculation unit is used to extract the quality root mean square error and loss stability index from the historical connection records of each template in the candidate connection template group, and calculate the comprehensive performance score. The template selection unit is used to sort the candidate template group in descending order according to the comprehensive performance score and select the optimal successor template; The reference curve acquisition unit is used to retrieve the set of historical loss curves associated with the optimal splicing template from the optical cable knowledge base; The collaborative analysis unit is used to perform trajectory-loss consistency checks on each curve in the historical loss curve set, remove abnormal curves with mismatch points, and form an optimized curve set. The baseline curve determination unit is used to select the curve with the smallest distortion as the baseline loss curve based on the historical stability of each curve in the optimization curve set. The parameter initialization unit is used to perform spatiotemporal registration of the optimal splicing template and the reference loss curve, and to set the initial splicing parameters.

10. The ODF-based optical cable splicing information management system according to claim 9, characterized in that, The collaborative analysis unit includes: The data pairing component is used to select the curve to be tested from the historical loss curve set and simultaneously obtain the trajectory parameter sequence of the corresponding time point in the optimal continuation template; The time alignment component is used to label the curve to be tested with collaborative time tags based on key event points in the trajectory parameter sequence; The mutation identification component is used to scan the time-stamped loss curve, detect whether the loss gradient in the neighborhood of each label exceeds the critical value, and mark the mutation interval. The collision detection component is used to compare the changing trend of trajectory parameters with the direction of the abrupt change in loss within the abrupt change range, and to calculate the collision intensity index. The collision marker component is used to record the start and end times of a collision interval when the collision intensity index exceeds a threshold. The curve reconstruction component is used to interpolate and generate smooth transition segments within the conflict interval based on historical conflict resolution cases. The curve replacement component is used to cover the original conflict zone with a smooth transition segment, generating an optimized loss curve. The integrity check component is used to verify the continuity of the optimized curve and filter out residual conflict points. The set of building components is used to integrate all validated curves into an optimized curve set.

Images