An intelligent optical fiber distribution system based on fault alarms

Abstract

The present invention relates to the technical field of optical fiber distribution management, and provides an intelligent optical fiber distribution system based on fault alarm, aiming to improve the fault detection and management capabilities of optical fiber networks; the system combines a U-Net model with multi-scale convolutional kernels, spatial transformation networks, and depthwise separable convolutions to provide more accurate and efficient feature extraction; in addition, the control module introduces a noise injection-based ε-greedy strategy on the basis of the Dueling-DQN and DDPG algorithms to optimize the optical fiber connection path and traffic allocation; the system not only enhances the recognition ability of complex fault patterns, but also realizes intelligent decision support in a dynamic environment; the present invention has significant practical value in optical fiber distribution management and provides users with more reliable network services.

Description

Technical Field

[0001] The present invention relates to the technical field of optical fiber distribution management, and particularly to an intelligent optical fiber distribution system based on fault alarm. Background Art

[0002] An intelligent optical fiber distribution system is a modern communication infrastructure management support system that integrates optical fiber communication technology and an intelligent network management system. Through automation and intelligent technology, it realizes the dynamic management, real-time monitoring, and efficient configuration of the optical fiber network. When traditional intelligent optical fiber distribution systems perform feature extraction, they often rely on predefined rules and can only extract some significant features, failing to deeply mine the complex patterns in the data. The data of the optical fiber network often has temporal and spatial dependencies. However, traditional intelligent optical fiber distribution systems cannot comprehensively capture the spatio-temporal features of network data. In the control module, when traditional intelligent optical fiber distribution systems face a complex and dynamic optical fiber network environment, traditional decision-making methods often cannot quickly find the optimal solution, resulting in low efficiency of resource allocation and traffic distribution. When facing sudden faults or traffic changes in the network, they often cannot make real-time adjustments, leading to a decrease in system performance or an expansion of network faults. Therefore, there is an urgent need for a more intelligent optical fiber distribution system to improve the real-time response ability, adaptability, and fault recovery ability of the intelligent optical fiber distribution system. Summary of the Invention

[0003] For an intelligent optical fiber distribution system based on fault alarm, a U-Net model combining multi-scale convolutional kernels, spatial transformation networks, and depthwise separable convolutions is proposed in terms of feature engineering. This technology can more effectively extract multi-level features in optical fiber data, not only enhancing the ability to capture high-frequency features but also improving the recognition rate of different fault modes, significantly improving the fault management efficiency of the optical fiber distribution system. In terms of the control module of the intelligent optical fiber distribution system, the present invention combines the ε-greedy strategy with noise injection, Dueling-DQN, and DDPG, which can enhance the stability and flexibility of control decisions in complex environments. When the intelligent optical fiber distribution system faces a dynamically changing network state, it can still effectively adjust the optical fiber connection path and traffic distribution without falling into a local optimal solution. By comprehensively considering environmental noise and uncertainty, the present invention optimizes the adaptability of the control strategy, thereby improving the overall performance and reliability of the intelligent optical fiber distribution system and providing a more intelligent and efficient solution for users.

[0004] The present invention provides an intelligent optical fiber distribution system based on fault alarm, which is characterized in that it includes a data acquisition module, a feature engineering module, and a control module;

[0005] The data acquisition module collects the connection status data, traffic load data, and fault alarm data of the optical fiber network as a data set;

[0006] Feature engineering module, establish a U-Net model, and based on the U-Net model, combine multi-scale convolutional kernels, spatial transformation networks, and depthwise separable convolutions to construct an MS-STS-DS-U-Net model. Use the MS-STS-DS-U-Net model to extract features from the dataset and generate an enhanced dataset; the MS-STS-DS-U-Net model includes an encoder and a decoder;

[0007] Control module, combine the ε-greedy strategy with noise injection, Dueling-DQN, and DDPG to construct the NI-ε-G-D3S algorithm. Analyze the enhanced dataset through the NI-ε-G-D3S algorithm to generate an optimal control decision, and issue instructions according to the optimal control decision to adjust the configuration of the optical fiber connection path and the traffic allocation.

[0008] Furthermore, the process of using the MS-STS-DS-U-Net model by the feature engineering module to extract features from the dataset and generate an enhanced dataset specifically includes the following steps:

[0009] Step S1: Data preprocessing: Perform spectral subtraction and feature alignment on the dataset to obtain a preprocessed dataset;

[0010] Step S2: Signal processing: Process the preprocessed dataset through the MS-STS-DS-U-Net model to separate and obtain single-event signals. The MS-STS-DS-U-Net model specifically includes a feature extraction unit, a compressed information unit, and a signal reconstruction unit;

[0011] Step S3: Enhanced dataset generation: Convert the single-event signal into a time-frequency image through short-time Fourier transform, extract the frequency and time features in the single-event signal to obtain time-frequency feature data; establish an AlexNet model, introduce a temperature parameter in the SoftMax layer of the AlexNet model to construct an improved AlexNet model, and use the improved AlexNet model to analyze the time-frequency feature data to generate an enhanced dataset.

[0012] Furthermore, the steps specifically executed by the feature extraction unit in Step S2 include:

[0013] Step M1: Input the preprocessed dataset into the multi-scale convolutional layer in the encoder, and extract fault signal features through three convolutional kernels of 3x3, 5x5, and 7x7 to capture the frequency and intensity changes at different scales in the fault signal features and obtain convolutional feature data;

[0014] Step M2: Process the convolutional feature data through a Sinc filter and a ReLU activation function to obtain non-linear feature data;

[0015] Step M3: Pass the convolutional feature data to the decoder through skip connections to retain high-resolution feature data.

[0016] Furthermore, the steps performed by the compression information unit in Step S2 specifically include: using depthwise separable convolution to compress the non-linear feature data, reducing the dimension, extracting fault information in the low-dimensional space to obtain compressed feature data. The formula used is as follows:

[0017] Formula for depth convolution operation:

[0018] ;

[0019] Among them, represents the non-linear feature data, represents the output result of the depth convolution operation, that is, "compressed non-linear feature data"; represents the channel, represents the th channel weight, represents the convolution operator, represents the part of the th channel of the non-linear feature data;

[0020] Formula for point convolution operation:

[0021] ;

[0022] Among them, represents the compressed feature data, represents the point convolution operation.

[0023] Furthermore, the steps performed by the reconstructed signal unit in Step S2 specifically include: passing the compressed feature data through deconvolution operation to obtain deconvolution feature data, then decoding the deconvolution feature data through a Sinc filter and ReLU activation function to obtain the decoded non-linear feature data, adding a spatial transformation network in the decoder to perform spatial transformation and correction on the decoded non-linear feature data, and combining with the high-resolution feature data to obtain a single event signal.

[0024] Furthermore, in Step S3, the process of using the improved AlexNet model to analyze the time-frequency feature data and generate an enhanced data set specifically includes the following steps:

[0025] Step S31: Establish an improved AlexNet model. The improved AlexNet model includes a convolutional layer, a fully connected layer, and a SoftMax layer. Input the time-frequency feature data into the improved AlexNet model, and use the convolutional layer to extract the spatial features in the time-frequency feature data, identify the frequency pattern and time dependence, and obtain high-level image feature data;

[0026] Step S32: The high-level image feature data is further processed through a fully connected layer to integrate spatial and temporal features, obtaining comprehensive feature data;

[0027] Step S33: Introduce a temperature parameter in the SoftMax layer, input the comprehensive feature data into the SoftMax layer to obtain a feature probability distribution, and combine the feature probability distribution to form an enhanced data set. The formula used is as follows:

[0028] ;

[0029] where, represents the index of a specific event category, represents the th unnormalized score of the th category, represents the weight factor, represents the SoftMax output adjusted by the temperature parameter and the weight factor, used to generate the probability distribution of each event category, represents the exponential function, represents the indices of all event categories, represents the normalization term.

[0030] Furthermore, in the control module, the process of analyzing the enhanced data set through the NI-ε-G-D3S algorithm to obtain the optimal control decision specifically includes the following steps:

[0031] Step B1: Initialize and establish a replay pool: Initialize the network weights of Dueling-DQN and DDPG, and establish an experience replay pool;

[0032] Step B2: Action selection: According to the enhanced data set, Dueling-DQN uses the noise-injected ε-greedy strategy to select a discrete action from the current state. The formula used is as follows:

[0033] ;

[0034] where, represents the current state, represents the discrete action, represents the adjustment after adding noise value, original value, represents the variance of the noise, represents the Gaussian noise term;

[0035] Step B3: Action parameterization: Parameterize the discrete action and use DDPG to generate a continuous action to form a comprehensive control decision;

[0036] Step B4: Data Sampling and Update: Randomly sample data from the experience replay pool, calculate the gradients of Dueling-DQN and DDPG, and update the network weights of Dueling-DQN and DDPG;

[0037] Step B5: Optimize Control Decision: Set the maximum number of loops, repeat the above steps until the maximum number of loops is reached, and finally output the optimal control decision.

[0038] Adopting the above solution, the beneficial effects of the present invention are as follows:

[0039] An intelligent optical fiber distribution system based on fault alarms provided by the present invention, in the feature engineering stage, through a U-Net model that combines multi-scale convolutional kernels, spatial transformation networks, and depthwise separable convolutions, can provide more accurate and efficient feature extraction when processing high-dimensional and complex optical fiber network data; multi-scale convolutional kernels can simultaneously process optical fiber data features at different scales, effectively capturing the diversity and complexity of various fault signals; through the spatial transformation network, the intelligent optical fiber distribution system can analyze and predict signals in the optical fiber network more accurately, improving the accuracy of fault detection and traffic optimization under different network configurations and changes in optical fiber connections; depthwise separable convolutions greatly reduce the computational amount and complexity of the intelligent optical fiber distribution system, helping to extract finer-grained features, improving the accuracy of network faults and traffic problems, and thus improving the accuracy and efficiency of fault diagnosis;

[0040] In the control module, the present invention combines the ε-greedy strategy with noise injection, Dueling-DQN, and DDPG strategies to achieve intelligent optimization of optical fiber connection paths and traffic allocation; this strategy enhances the decision-making ability of the intelligent optical fiber distribution system in a dynamic environment, enabling it to maintain efficient control decisions when facing uncertainties; by introducing the influence of environmental noise, this method effectively improves the adaptability of the system to real-time data, thereby optimizing the overall network performance and reducing the risk of faults;

[0041] In summary, through the combination of the above technologies, the present invention not only improves the fault detection and control capabilities of the intelligent optical fiber distribution system, but also provides a more intelligent solution for network management; enabling the system to operate efficiently and stably in the face of complex and changing network environments, and providing more reliable services for users. Brief Description of the Drawings

[0042] Figure 1 It is a schematic diagram of the modules of an intelligent optical fiber distribution system based on fault alarms proposed by the present invention;

[0043] Figure 2Flow schematic diagram of the feature engineering module proposed in the second embodiment;

[0044] Figure 3 Flow schematic diagram of the NI-ε-G-D3S algorithm proposed in the tenth embodiment. Detailed implementation manners

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0046] Embodiment 1, according to Figure 1 , the present invention provides an intelligent optical fiber distribution system based on fault alarms, and the system includes a data acquisition module, a feature engineering module, and a control module;

[0047] The data acquisition module acquires connection status data, traffic load data, and fault alarm data of the optical fiber network as a data set;

[0048] The feature engineering module establishes a U-Net model, and on the basis of the U-Net model, combines a multi-scale convolutional kernel, a spatial transformation network, and a depthwise separable convolution to construct an MS-STS-DS-U-Net model, and uses the MS-STS-DS-U-Net model to extract features from the data set to generate an enhanced data set; the MS-STS-DS-U-Net model includes an encoder and a decoder;

[0049] The control module combines the ε-greedy strategy with noise injection, Dueling-DQN, and DDPG to construct an NI-ε-G-D3S algorithm, analyzes the enhanced data set through the NI-ε-G-D3S algorithm to generate an optimal control decision, and issues an instruction according to the optimal control decision to adjust the configuration of the optical fiber connection path and the traffic distribution.

[0050] Embodiment 2, according to Figure 2 , this embodiment is based on Embodiment 1. In this embodiment, the process of using the MS-STS-DS-U-Net model by the feature engineering module to extract features from the data set to generate an enhanced data set specifically includes the following steps:

[0051] Step S1: Data preprocessing: Perform spectral subtraction and feature alignment on the data set to obtain a preprocessed data set;

[0052] Step S2: Signal Processing: Process the preprocessed dataset through the MS-STS-DS-U-Net model to separate and obtain single-event signals. The MS-STS-DS-U-Net model specifically includes a feature extraction unit, a compressed information unit, and a signal reconstruction unit;

[0053] Step S3: Augmented Dataset Generation: Convert the single-event signals into time-frequency images through short-time Fourier transform, extract the frequency and time features in the single-event signals to obtain time-frequency feature data; establish an AlexNet model, introduce a temperature parameter into the SoftMax layer of the AlexNet model to construct an improved AlexNet model, and use the improved AlexNet model to analyze the time-frequency feature data to generate an augmented dataset.

[0054] Example 3. This example is based on Example 1. In this example, the process of the feature engineering module generating an augmented dataset specifically includes the following steps:

[0055] Step E1: Data Preprocessing: Perform spectral subtraction and feature alignment on the dataset to obtain a preprocessed dataset;

[0056] Step E2: Signal Processing: Process the preprocessed dataset through a U-Net model to separate and obtain single-event signals;

[0057] Step E3: Augmented Dataset Generation: Convert the single-event signals into time-frequency images through short-time Fourier transform, extract the frequency and time features in the single-event signals to obtain time-frequency feature data; establish an AlexNet model, introduce a temperature parameter into the SoftMax layer of the AlexNet model to construct an improved AlexNet model, and use the improved AlexNet model to analyze the time-frequency feature data to generate an augmented dataset.

[0058] Example 4. This example is based on Example 2. In this example, the steps specifically executed by the feature extraction unit in Step S2 include:

[0059] Step M1: Input the preprocessed dataset into the multi-scale convolutional layer in the encoder, extract the fault signal features through three convolutional kernels of 3x3, 5x5, and 7x7, capture the frequency and intensity changes at different scales in the fault signal features to obtain convolutional feature data;

[0060] Step M2: Process the convolutional feature data through a Sinc filter and a ReLU activation function to obtain non-linear feature data;

[0061] Step M3: Transmit the convolutional feature data to the decoder through skip connections to retain the high-resolution feature data.

[0062] Embodiment 5. This embodiment is based on Embodiment 4. In this embodiment, the steps performed by the compression information unit in step S2 specifically include: using depthwise separable convolution to compress the non-linear feature data, reducing the dimension, extracting fault information in the low-dimensional space to obtain compressed feature data. The formula used is as follows:

[0063] Formula for depth convolution operation:

[0064] ;

[0065] Among them, represents the non-linear feature data, represents the output result of the depth convolution operation, that is, "compressed non-linear feature data"; represents the channel, represents the th channel weight, represents the convolution operator, represents the part of the non-linear feature data in the th channel;

[0066] Formula for point convolution operation:

[0067] ;

[0068] Among them, represents the compressed feature data, represents the point convolution operation.

[0069] Embodiment 6. This embodiment is based on Embodiment 5. In this embodiment, the steps performed by the reconstructed signal unit in step S2 specifically include: passing the compressed feature data through a transposed convolution operation to obtain transposed convolution feature data, then decoding the transposed convolution feature data through a Sinc filter and a ReLU activation function to obtain the decoded non-linear feature data, adding a spatial transformation network in the decoder to perform spatial transformation and correction on the decoded non-linear feature data, and combining with high-resolution feature data to obtain a single event signal.

[0070] Embodiment 7. This embodiment is based on Embodiment 5. In this embodiment, the steps performed by the reconstructed signal unit in step S2 specifically include: passing the compressed feature data through a transposed convolution operation to obtain transposed convolution feature data, then decoding the transposed convolution feature data through a Sinc filter and a ReLU activation function to obtain the decoded non-linear feature data, and combining with high-resolution feature data to obtain a single event signal.

[0071] Embodiment 8. This embodiment is based on Embodiment 6. In this embodiment, in step S3, the process of using the improved AlexNet model to analyze the time-frequency feature data and generate an enhanced data set specifically includes the following steps:

[0072] Step S31: Establish an improved AlexNet model. The improved AlexNet model includes a convolutional layer, a fully connected layer, and a SoftMax layer. Input the time-frequency feature data into the improved AlexNet model, and use the convolutional layer to extract the spatial features in the time-frequency feature data, identify the frequency patterns and time dependencies, and obtain high-level image feature data;

[0073] Step S32: The high-level image feature data is further processed through the fully connected layer to integrate the spatial features and time features, and obtain comprehensive feature data;

[0074] Step S33: Introduce a temperature parameter in the SoftMax layer. Input the comprehensive feature data into the SoftMax layer to obtain a feature probability distribution, and combine the feature probability distribution to form an enhanced data set. The formula used is as follows:

[0075] ;

[0076] where, represents the index of a specific event category, represents the unnormalized score of the th category, represents the weight factor, represents the temperature parameter, represents the SoftMax output adjusted by the temperature parameter and the weight factor, which is used to generate the probability distribution of each event category, represents the exponential function, represents the index of all event categories, represents the normalization term.

[0077] Example 9. This example is based on Example 6. In this example, in step S3, the process of using the improved AlexNet model to analyze the time-frequency feature data and generate an enhanced data set specifically includes the following steps:

[0078] Step N1: Establish an improved AlexNet model. The improved AlexNet model includes a convolutional layer, a fully connected layer, and a SoftMax layer. Input the time-frequency feature data into the improved AlexNet model, and use the convolutional layer to extract the spatial features in the time-frequency feature data, identify the frequency patterns and time dependencies, and obtain high-level image feature data;

[0079] Step N2: The high-level image feature data is further processed through the fully connected layer to integrate the spatial features and time features, and obtain comprehensive feature data;

[0080] Step N3: Input the comprehensive feature data into the SoftMax layer to obtain the feature probability distribution, and combine the feature probability distribution to form an enhanced data set.

[0081] Embodiment Ten. According to Figure 3 , this embodiment is based on Embodiment Eight. In this embodiment, in the control module, the process of analyzing the enhanced data set through the NI-ε-G-D3S algorithm to obtain the optimal control decision specifically includes the following steps:

[0082] Step B1: Initialize and establish a replay pool: Initialize the network weights of Dueling-DQN and DDPG, and establish an experience replay pool;

[0083] Step B2: Action selection: According to the enhanced data set, Dueling-DQN uses the noise-injected ε-greedy strategy to select discrete actions from the current state. The formula used is as follows:

[0084] ;

[0085] Where represents the current state, represents the discrete action, represents the adjusted value after adding noise, the original value, represents the variance of the noise, represents the Gaussian noise term;

[0086] Step B3: Action parameterization: Parameterize the discrete action and use DDPG to generate a continuous action to form a comprehensive control decision;

[0087] Step B4: Data sampling and update: Randomly sample data from the experience replay pool, calculate the gradients of Dueling-DQN and DDPG, and update the network weights of Dueling-DQN and DDPG;

[0088] Step B5: Optimize the control decision: Set the maximum number of loops, repeat the above steps until the maximum number of loops is reached, and finally output the optimal control decision;

[0089] The maximum number of loops is 500 times.

[0090] The above describes the present invention and its implementation manners. This description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto; generally speaking, if those of ordinary skill in the art are inspired by it and design similar structural manners and embodiments without creative efforts without departing from the purpose of the present invention's creation, they should all fall within the protection scope of the present invention.

Claims

1. An intelligent optical fiber wiring system based on fault alarm, comprising a data acquisition module, wherein the data acquisition module collects connection status data, flow load data and fault alarm data of an optical fiber network as a data set; characterized in that: The system also includes a feature engineering module and a control module; The feature engineering module establishes a U-Net model, combines multi-scale convolution kernels, spatial transformation networks and depth-separable convolutions on the basis of the U-Net model, constructs an MS-STS-DS-U-Net model, uses the MS-STS-DS-U-Net model to extract features from the data set, and generates an enhanced data set; The MS-STS-DS-U-Net model includes an encoder and a decoder; The control module combines the ε-greedy strategy of noise injection, Dueling-DQN and DDPG to construct the NI-ε-G-D3S algorithm, analyzes the enhanced data set through the NI-ε-G-D3S algorithm, generates the optimal control decision, issues instructions according to the optimal control decision, and adjusts the configuration and flow distribution of the optical fiber connection path; The feature engineering module generates an enhanced data set, which specifically includes the following steps: Step S1: Data preprocessing: performing spectrum subtraction and feature alignment on the data set to obtain a preprocessed data set; Step S2: signal processing: the preprocessing data set is processed by the MS-STS-DS-U-Net model to separate and obtain a single event signal, wherein the MS-STS-DS-U-Net model specifically includes a feature extraction unit, an information compression unit, and a signal reconstruction unit; Step S3: Enhanced data set generation: Convert a single event signal into a time-frequency image through short-time Fourier transform, extract the frequency and time features in the single event signal, and obtain time-frequency feature data; establish an AlexNet model, introduce temperature parameters in the SoftMax layer of the AlexNet model, construct an improved AlexNet model, use the improved AlexNet model to analyze the time-frequency feature data, and generate an enhanced data set.

2. The intelligent optical fiber wiring system based on fault alarm according to claim 1, characterized in that: The steps of extracting feature units specifically include: Step M1: Input the preprocessed data set into the multi-scale convolution layer in the encoder, extract the fault signal features through three convolution kernels of 3x3, 5x5, and 7x7, capture the frequency and intensity changes of different scales in the fault signal features, and obtain the convolution feature data; Step M2: Process the convolution feature data through a Sinc filter and a ReLU activation function to obtain nonlinear feature data; Step M3: Pass the convolutional feature data to the decoder through the skip connection, retaining the high-resolution feature data.

3. The intelligent optical fiber wiring system based on fault alarm according to claim 2, characterized in that: The steps performed by the compression information unit specifically include: using depth-separable convolution to compress nonlinear feature data, reduce the dimension, extract fault information in the low-dimensional space, and obtain compressed feature data.

4. The intelligent optical fiber wiring system based on fault alarm according to claim 3, characterized in that: The steps of reconstructing the signal unit specifically include: performing a deconvolution operation on the compressed feature data to obtain deconvolution feature data, decoding the deconvolution feature data through a Sinc filter and a ReLU activation function to obtain decoded nonlinear feature data, adding a spatial transformation network to the decoder, performing spatial transformation and correction on the decoded nonlinear feature data, and combining the high-resolution feature data to obtain a single event signal.

5. The intelligent optical fiber wiring system based on fault alarm according to claim 4, characterized in that: In step S3, the process of using the improved AlexNet model to analyze the time-frequency feature data and generate an enhanced data set specifically includes the following steps: Step S31: establishing an improved AlexNet model, the improved AlexNet model includes a convolutional layer, a fully connected layer and a SoftMax layer, inputting the time-frequency feature data into the improved AlexNet model, using the convolutional layer to extract spatial features in the time-frequency feature data, identifying frequency patterns and time dependencies, and obtaining high-level image feature data; Step S32: The high-level image feature data is further processed through a fully connected layer to integrate spatial features and temporal features to obtain comprehensive feature data; Step S33: Introduce the temperature parameter in the SoftMax layer, input the comprehensive feature data into the SoftMax layer, obtain the feature probability distribution, and combine the feature probability distribution to form an enhanced data set.

6. The intelligent optical fiber wiring system based on fault alarm according to claim 1, characterized in that: The process of obtaining the optimal control decision by the control module specifically includes the following steps: Step B1: Initialize and establish a replay pool: Initialize the network weights of Dueling-DQN and DDPG, and establish an experience replay pool; Step B2: Action selection: Based on the augmented dataset, Dueling-DQN uses the noise-injected ε-greedy strategy to select discrete actions from the current state; Step B3: Action parameterization: Parameterize discrete actions and use DDPG to generate continuous actions to form comprehensive control decisions; Step B4: Data sampling and updating: Randomly sample data from the experience replay pool, calculate the gradients of Dueling-DQN and DDPG, and update the network weights of Dueling-DQN and DDPG; Step B5: Optimize control decision: Set the maximum number of cycles, repeat the above steps until the maximum number of cycles is reached, and finally output the optimal control decision.

Images