Optical fiber fault diagnosis method and system based on SMC-BiLSTM network model

Abstract

The invention discloses an optical fiber fault diagnosis method and system based on an SMC-BiLSTM network model, and the method comprises the following steps: measuring an optical fiber through an optical time domain reflectometer, and obtaining original optical time domain reflectometer data corresponding to different fault types; preprocessing the original optical time domain reflectometer data to obtain preprocessed optical time domain reflectometer data; the preprocessed optical time domain reflectometer data is input into an SMC-BiLSTM network model for training, and a trained SMC-BiLSTM network model is obtained; using the trained SMC-BiLSTM network model to diagnose the optical time domain reflectometer data to be identified, and obtaining the type of the optical fiber fault and the diagnosis result; according to the method, the optical time domain reflectometer data is acquired and preprocessed, and training and identification are performed based on the SMC-BiLSTM network model, so that intelligent diagnosis of optical fiber faults is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optical fiber fault identification, and more particularly to an optical fiber fault diagnosis method and system based on an SMC-BiLSTM network model. BACKGROUND

[0002] For decades, optical time domain reflectometry (OTDR) technology has been the cornerstone in the field of optical fiber monitoring and fault analysis. Traditional methods, such as the two-point method combined with least square fitting, have been widely used to estimate the event location and loss in the OTDR trace by calculating the best fitting line between two markers. These methods, although simple, have limited accuracy in noisy environments, making them less suitable for modern optical networks that require increasing reliability.

[0003] To address these limitations, researchers have proposed various advanced techniques. Liu et al. introduced the Gabor Series Representation combined with the Minimum Description Length criterion (GSR / MDL) method, which was later improved by the R1MSDE method, achieving higher event detection accuracy using the theory of matching subspace detection. Man et al. developed an event detection method based on short-time Fourier transform (STFT), combined with exponential windowing and binary signal detection theory, enhancing the ability to detect distant events in OTDR and improving computational efficiency. In addition, methods based on wavelet transform (WT) show promise due to their scalable spatial resolution. However, the computational complexity of wavelet transform often hinders its real-time application in optical fiber monitoring systems. Subsequently, Kong et al. proposed a hybrid method combining correlation matching with short-time Fourier transform (STFT), but these methods still require a large number of numerical values and are difficult to adapt to low signal-to-noise ratio (SNR) environments.

[0004] In recent years, machine learning techniques have demonstrated powerful capabilities in OTDR event analysis. Abdelli et al. showcased the wide application of machine learning in laser failure mode detection, lifetime prediction, and reliability enhancement. They also demonstrated the potential of convolutional neural networks (CNNs) to detect and characterize reflection events in noisy environments. Furthermore, Abdelli et al. used long-short-term memory (LSTM) networks to detect, locate, and estimate the reflectivity of fiber optic reflection faults (events), including connectors and mechanical joints. In addition, Zhang et al. demonstrated the robustness of deep convolutional neural networks combined with novel training methods in noisy environments, providing theoretical support for handling noise in OTDR signals. Bidirectional long-short-term memory (BiLSTM) networks have been used to address time-dependent issues in OTDR trajectories for fault detection and localization; the combination of CNNs and BiLSTMs has proven effective in improving the accuracy and robustness of fault diagnosis tasks.

[0005] Therefore, the generalization ability and robustness of these models can deteriorate significantly when tested with new, unknown data, including real-world induced failures of various modes, such as fiber bending events at different bending radius values. Furthermore, while these methods can distinguish non-reflective events from other events, they cannot easily differentiate failures caused by poor splicing or fiber connectors. However, most current methods focus on single-scale analysis or rely on fixed feature extraction techniques, making it difficult to fully capture the multi-scale characteristics of OTDR signals. Summary of the Invention

[0006] The purpose of this invention is to provide a fiber optic fault diagnosis method and system based on the SMC-BiLSTM network model. By collecting and preprocessing optical time-domain reflectometer data, and training and identifying based on the SMC-BiLSTM network model, intelligent diagnosis of fiber optic faults can be achieved.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A fiber optic fault diagnosis method based on the SMC-BiLSTM network model includes the following steps: The optical fiber was measured using an optical time domain reflectometer to obtain the original optical time domain reflectometer data corresponding to different fault types; The original optical time domain reflectometer data is preprocessed to obtain preprocessed optical time domain reflectometer data. The preprocessed optical time domain reflectometer data is input into the SMC-BiLSTM network model for training, and the trained SMC-BiLSTM network model is obtained. The trained SMC-BiLSTM network model is used for diagnosing the optical time domain reflectometer data to be identified, so as to obtain the type and diagnosis result of the optical fiber fault.

[0008] Further, the original optical time domain reflectometer data is specifically backscattering signal data collected by an optical time domain reflectometer; and the original optical time domain reflectometer data includes six different types of optical fiber faults: fiber cutting, eavesdropping, splicing, faulty connector, bending and physical contact (PC) connector. The optical time domain reflectometer data contains 125,832 records, the pulse width of the optical time domain reflectometer configuration is 10 ns, the wavelength is 1650 nm, and the sampling time is 1 ns. The total length of the test optical cable is about 14 kilometers.

[0009] Further, the architecture of the SMC-BiLSTM network model includes an input layer, a convolutional neural network layer, a bidirectional long short-term memory network layer and a deep neural network layer. The input layer is configured to receive the preprocessed optical time domain reflectometer data and convert the data into an input tensor. The convolutional neural network layer is configured to perform multi-scale local feature extraction and spatial dimension reduction on the input tensor to generate a time series feature map. The bidirectional long short-term memory network layer is configured to model the bidirectional time sequence dependence of the time series in the time series feature map through a parameter sharing mechanism to generate a feature vector that fuses global context. The deep neural network layer is configured to perform high-level feature abstraction and fault classification on the feature vector.

[0010] Further, the convolutional neural network layer includes a convolutional layer and a pooling layer. The convolutional layer is configured to slide a multi-scale convolutional kernel over the input tensor to extract local spatial features, and introduce a non-linear transformation through a ReLU activation function to generate a time series feature map. The pooling layer is configured to downsample the time series feature map to reduce the dimension of the time series feature map while retaining important features.

[0011] Further, the bidirectional long short-term memory network layer includes a forward LSTM network and a backward LSTM network, which are configured to capture the forward and backward dependencies in the time series feature map to generate a feature vector that fuses global context.

[0012] Further, the deep neural network layer includes a fully connected layer, a dropout layer and an output layer. The fully connected layer is configured to map the feature vector to a high-dimensional space, extract non-linear combination features layer by layer, and introduce a ReLU activation function to enhance the model expression ability. The Dropout layer is used for preventing overfitting of the SMC-BiLSTM network model by randomly discarding a part of neurons. The output layer is used for converting the feature vector into a probability distribution vector of the fault type through a Softmax function.

[0013] Further, the parameters of the bidirectional long short-term memory network layer are consistent when processing all CNN branch outputs, and the total amount of model parameters is reduced through a parameter sharing mechanism.

[0014] Further, the number of the CNN branches is four, and the convolution kernel scales of the branches are sequentially increased to capture spatial features of different granularities.

[0015] Further, the process of fusing the feature vector of the global context is to concatenate the feature dimensions to form the feature vector of the fused global context.

[0016] The application also provides a system of an optical fiber fault diagnosis method based on an SMC-BiLSTM network model, comprising: A data acquisition module is configured to measure an optical fiber by using an optical time domain reflectometer to obtain original optical time domain reflectometer data corresponding to different fault types. A multi-scale feature extraction module is configured to preprocess the original optical time domain reflectometer data to obtain preprocessed optical time domain reflectometer data. A model training module is configured to input the preprocessed optical time domain reflectometer data into an SMC-BiLSTM network model for training to obtain a trained SMC-BiLSTM network model. A fault recognition module is configured to use the trained SMC-BiLSTM network model to diagnose optical time domain reflectometer data to be recognized to obtain a type and a diagnosis result of an optical fiber fault.

[0017] According to the embodiments of the application, the following technical effects are achieved: The application realizes efficient recognition and diagnosis of optical fiber faults by obtaining and preprocessing original optical time domain reflectometer (OTDR) data and combining a shared multi-scale convolution bidirectional long short-term memory network (SMC-BiLSTM) model. Further, the model parameter amount is significantly reduced, and the deployment capability of the model in a resource-limited environment is enhanced by extracting local features through a multi-scale CNN and using a shared BiLSTM layer to unify modeling of time sequence dependency. Meanwhile, the accuracy and robustness of fault recognition are effectively improved through bidirectional time sequence feature fusion and deep neural network classification, and the application is suitable for real-time and high-precision optical fiber fault monitoring and diagnosis scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative effort based on the provided drawings.

[0019] The SMC-BiLSTM network model-based optical fiber fault diagnosis method and system of the present application will be further described below in combination with the accompanying drawings. Figure 1 is a schematic diagram of the overall process of the SMC-BiLSTM network model-based optical fiber fault diagnosis method of embodiment 1 in the present application. Figure 2 is a network architecture diagram of the SMC-BiLSTM network model of embodiment 1 in the present application. Figure 3 is a recognition result diagram of the SMC-BiLSTM network model of embodiment 2 in the present application. Figure 4 is a ROC curve diagram of the SMC-BiLSTM network model of embodiment 2 in the present application. Figure 5 is a visualization diagram of feature learning of the SMC-BiLSTM network model of embodiment 2 in the present application. DETAILED DESCRIPTION

[0020] The specific embodiments of the present application will be further described in detail below in combination with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present application, but not to limit the scope of the present application.

[0021] In order to better understand the purpose, structure and function of the present application, the present application will be further described in detail below in combination with the accompanying drawings.

[0022] Embodiment 1 The present application provides an SMC-BiLSTM network model-based optical fiber fault diagnosis method, as shown in Figure 1 , comprising the following steps: Measuring the optical fiber using an optical time domain reflectometer to obtain original optical time domain reflectometer data corresponding to different fault types; Pretreating the original optical time domain reflectometer data to obtain pretreated optical time domain reflectometer data; In this embodiment, a small number of class fault samples (such as eavesdropping and bending) are subjected to time series data disturbance (such as adding Gaussian noise, time distortion, and random scaling) to generate synthetic samples; Layered sampling is used: when dividing training / validation / testing sets, stratified sampling is performed according to the fault type proportion to ensure consistent distribution of each category.

[0023] The preprocessing comprises the following steps: (a) denoising: using wavelet threshold denoising method (such as Symlets wavelet basis) to eliminate Gaussian noise in the original OTDR data; (b) normalization: Min-Max normalization is performed on the denoised data, and the amplitude is mapped to the interval [0, 1]; (c) segmentation: the continuous OTDR curve is intercepted by a sliding window with a fixed length (such as 512 points) to generate time series samples; (d) labeling: according to the fault position information (such as reflection peak, step drop point), the sample is assigned a fault type label; it helps to improve data quality, eliminate device difference interference, and enhance model generalization ability.

[0024] The preprocessed optical time domain reflectometer data is input into the SMC-BiLSTM network model for training, and a trained SMC-BiLSTM network model is obtained; The trained SMC-BiLSTM network model is used to diagnose the optical time domain reflectometer data to be identified, and the type and diagnosis result of the optical fiber fault are obtained.

[0025] The diagnosis result includes: fault type (optical fiber cutting / eavesdropping / splicing / faulty connector / bending / PC connector), fault position (converted to physical distance based on OTDR time sequence index), fault severity (such as reflectivity, loss value) and confidence probability; it can provide a structured diagnosis report to support operation and maintenance decisions (such as prioritizing high-confidence faults).

[0026] The original optical time domain reflectometer data is specifically the backscattering signal data collected by the optical time domain reflectometer.

[0027] As shown in Figure 2 The architecture of the SMC-BiLSTM network model comprises: an input layer, a convolutional neural network layer, a bidirectional long short-term memory network layer and a deep neural network layer; The input layer is used to receive the preprocessed optical time domain reflectometer data and convert it into an input tensor; The convolutional neural network layer is used to perform multi-scale local feature extraction and spatial dimension reduction on the input tensor to generate a time series feature map; The bidirectional long short-term memory network layer is used to model the bidirectional time sequence dependence of the time series in the time series feature map through a parameter sharing mechanism to generate a feature vector that fuses global context; The deep neural network layer is used to perform high-level feature abstraction and fault classification on the feature vector.

[0028] The convolutional neural network layer comprises a convolutional layer and a pooling layer. The convolutional layer is configured to slide over an input tensor through a multi-scale convolutional kernel, extract local spatial features, and introduce a non-linear transformation through a ReLU activation function to generate a time series feature map. The pooling layer is configured to down-sample the time series feature map to reduce the dimension of the time series feature map while retaining important features.

[0029] The bidirectional long short-term memory network layer comprises a forward LSTM network and a backward LSTM network, which are configured to capture the forward and backward dependencies in the time series feature map to generate a feature vector that fuses global context.

[0030] The deep neural network layer comprises a Batch Normalization layer, a fully connected layer, a dropout layer, and an output layer.

[0031] The fully connected layer is configured to map the feature vector to a high-dimensional space, extract non-linear combination features layer by layer, and introduce a ReLU activation function to enhance the model expression capability. The dropout layer is configured to prevent overfitting of the SMC-BiLSTM network model by randomly discarding a portion of neurons. The output layer is configured to convert the feature vector into a probability distribution vector of fault types through a Softmax function.

[0032] The parameters of the bidirectional long short-term memory network layer are consistent when processing all CNN branch outputs, and the total amount of model parameters is reduced through a parameter sharing mechanism.

[0033] The number of CNN branches is four, and the convolutional kernel scales of each branch are sequentially increased to capture spatial features of different granularities.

[0034] In this embodiment, the convolutional kernel scales of the four CNN branches are 3x1, 5x1, 7x1, and 9x1, respectively, and the step size is 1.

[0035] The feature maps output by all CNN branches share the same set of BiLSTM weight parameters, enabling the model to capture multi-scale features while reducing the number of parameters by more than 40%, significantly reducing computational complexity and improving real-time performance. Multi-scale convolutional kernels can cover features of different granularities (such as sharp reflection peaks and slowly varying loss regions), improving the recognition accuracy of complex faults (such as micro-bending and joint loss).

[0036] Example 2 This embodiment also provides another design scheme for convolutional neural network (CNN) layers and bidirectional long short-term memory (BiLSTM) layers: In addition to the four CNN branches (3×1, 5×1, 7×1, 9×1) in Example 1, a dilated convolution branch (e.g., 3×1, dilation rate = 2) is added to expand the receptive field to capture long-range dependent features, thereby improving the ability to identify gradual faults (e.g., microbending loss).

[0037] Furthermore, a residual block is introduced into the CNN branch to address the vanishing gradient problem in deep networks, ensuring that feature information is not lost during multi-level propagation. The function expression is as follows: ; in, x The input feature map (the output of the previous layer network or the initial input); Conv( x ):right x Perform convolution operations; ReLU(·) is the ReLU activation function (which truncates negative values ​​of the convolution output to 0, retains positive values, and introduces non-linearity). A self-attention layer is added after the BiLSTM output to dynamically weight the features at key time steps, thereby improving the model's sensitivity to fault abrupt changes (such as reflection peaks). The function expression is as follows: ; in, Q For query matrix; K The key matrix; V It is a value matrix; The dimension of the key vector; softmax(·) is the softmax normalization; Initialize the forget gate bias of BiLSTM to 1 (refer to Jozefowicz et al.) to alleviate gradient vanishing and accelerate the convergence of long sequence training.

[0038] In addition to fault classification, a new regression branch is added to synchronously output fault location (physical distance) and severity (loss value in dB). The loss function is designed as follows: ; wherein, is the total loss function, which integrates the loss of three tasks: classification, localization, and severity; is the classification loss, which measures the prediction error of fault type, such as cross-entropy loss; is the location regression loss, which measures the prediction error of physical distance, such as mean square error MSE ; is the severity regression loss, which measures the prediction error of loss value (dB), such as MSE or MAE; , , are the task weight coefficients, respectively, which are determined by validation set tuning to balance the contribution of different tasks, for example, increasing by validation set tuning.

[0039] In the training phase, adversarial samples (such as FGSM method to generate perturbation data) are injected to improve the generalization ability of the model in low signal-to-noise ratio (SNR < 15 dB) environment.

[0040] Embodiment 3 The application also provides a performance evaluation process of the SMC-BiLSTM network model in Embodiment 1: Before training the model, stratified random sampling is performed based on the dataset samples, and the samples are divided into training set, validation set and test set in the ratio of 6:2:2.

[0041] Optimizer: Adam; initial learning rate: 0.001; Batch size: 512; number of training rounds: 500; Early Stopping: based on the loss of the validation set to avoid overfitting; activation function: ReLU (F1 score is the highest) The original optical time domain reflectometer data is specifically: backscattering signal data collected by an optical time domain reflectometer; wherein the original optical time domain reflectometer data corresponds to six different types of optical fiber faults: optical fiber cutting, eavesdropping, splicing, faulty connector, bending and physical contact (PC) connector. The optical time domain reflectometer data contains 125,832 records, the pulse width of the optical time domain reflectometer configuration is 10 ns, the wavelength is 1650 nm, and the sampling time is 1 ns. The total length of the test optical cable is about 14 kilometers.

[0042] During the training process, in order to prevent the problems of data imbalance and overfitting, weighted cross-entropy loss, dynamic learning rate adjustment and data enhancement are used in turn.

[0043] The weighted cross-entropy loss is: the weight is assigned according to the proportion of fault samples (for example, the weight is high when the bending fault sample is small), and the expression is: ; Wherein is the category weight.

[0044] The dynamic learning rate adjustment is specifically: the cosine annealing learning rate scheduling: the initial learning rate is 0.001, and is decayed by the cosine function every 10 epochs; local optimum is avoided, and the convergence stability is improved.

[0045] The data enhancement is specifically: time series data enhancement: Gaussian noise (SNR=20dB) is randomly added to the OTDR curve, and the time axis is scaled (±5%), which expands the training data of small sample fault types (such as eavesdropping and physical contact connector).

[0046] After training, various traditional evaluation indexes are used to comprehensively evaluate the abnormal diagnosis performance of the model. Here, TP (true) represents correct prediction, TN (false) represents incorrect prediction, FP represents incorrect prediction of the correct category, and FN represents incorrect prediction of the wrong category. Based on the above definition, the accuracy, precision, recall and F1 score are used as classification evaluation indexes to evaluate the experimental results of the model, the ROC curve is obtained by plotting the relationship between TP and TN, and is visualized. As shown in Figure 4 and Figure 5 The optical fiber diagnosis result of the SMC-BiLSTM network model is shown in Figure 3 .

[0047] Based on the above definition, the accuracy, precision, recall and F1 score are used as classification evaluation indexes to evaluate the experimental results of the model. The accuracy Acc is the proportion of correctly classified samples to the total number of samples, indicating the overall correctness of the model classification.

[0048] ; The precision Prc is the proportion of true positive samples in all positive samples predicted by the model. It reflects the ability of the model to reduce false positives.

[0049] ; The recall Rc is the proportion of actual positive samples correctly predicted by the model. It represents the ability of the model to identify all positive cases, also known as sensitivity.

[0050] ; The F1 score is the harmonic mean of precision and recall, providing a balanced evaluation index considering false positives (FP) and false negatives (FN).

[0051] ; Latency: Optical fiber is a latency-sensitive network, and less latency is required when classifying the attack type. In the proposed system, the latency metric is the average time of each ML classifier used for testing, which indicates the computational efficiency in terms of processing time. Through this evaluation, the present invention aims to identify the classifier that achieves the best balance between accuracy and computational efficiency to ensure effective fault detection while minimizing processing time.

[0052] Embodiment 3 The present invention also provides an optical fiber fault diagnosis system applied to Embodiment 1, comprising: A data acquisition module is used to measure the optical fiber using an optical time domain reflectometer to obtain original optical time domain reflectometer data corresponding to different fault types; by acquiring original optical time domain reflectometer (OTDR) data; by obtaining original optical time domain reflectometer data without any modification, the physical characteristics of the optical fiber (such as backscattering power, Fresnel reflection, etc.) are converted into digital signals that can be processed by a computer, laying a solid and reliable data foundation for all subsequent intelligent analysis and model training, ensuring the source accuracy and effectiveness of the diagnosis results; In this embodiment, the data acquisition module receives OTDR data in real time with a sliding window (window length 512 points, step length 128 points).

[0053] A data processing module is used to preprocess the original optical time domain reflectometer data to obtain preprocessed optical time domain reflectometer data; by a series of preprocessing operations such as denoising, normalization, feature extraction, and event point labeling, the interference caused by measurement noise and equipment differences is effectively eliminated, the key features related to faults are highlighted, and unstructured curve data is converted into structured information that is easy for models to understand and learn. This not only greatly improves the efficiency and convergence speed of subsequent model training, but also significantly enhances the robustness and generalization ability of the final fault recognition model; A model training module is used to input the preprocessed optical time domain reflectometer data into an SMC-BiLSTM network model for training to obtain a trained SMC-BiLSTM network model; the model can autonomously learn the deep mapping relationship between the complex spatiotemporal patterns contained in the OTDR curve and various fault types. In particular, the BiLSTM structure can capture both forward and backward dependency information of the curve, while the SMC method helps to handle the uncertainty of the model, ultimately generating an expert-level diagnostic model that can accurately and automatically distinguish different faults (such as break points, macro bends, excessive fusion point loss, etc.); The fault identification module is used for diagnosing the optical time domain reflectometer data to be identified by using the trained SMC-BiLSTM network model, and obtaining the type and diagnosis result of the optical fiber fault; wherein the fault identification module outputs the confidence threshold (such as >0.9) to trigger the alarm, which can reduce false positives; the trained intelligent model is converted into actual productivity, and rapid, automatic and accurate diagnosis of the optical fiber fault is realized. The module receives the OTDR data to be identified, and after the same preprocessing process as in the training, inputs the OTDR data to the constructed SMC-BiLSTM model, and the model can quickly output a structured diagnosis result, which clearly indicates the type, position and severity of the fault. This completely replaces the cumbersome process of traditional curve interpretation relying on manual experience, shortens the fault positioning time from hours to seconds, greatly improves the operation and maintenance efficiency, reduces the labor cost, and ensures the consistency and objectivity of the diagnosis result.

[0054] The above description of the disclosed embodiments enables one skilled in the art to make or use the application. Numerous modifications to these embodiments will be apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for diagnosing optical fiber faults based on an SMC-BiLSTM network model, characterized in that, The method comprises the following steps: measuring the optical fiber by using an optical time domain reflectometer to obtain original optical time domain reflectometer data corresponding to different fault types; preprocessing the original optical time domain reflectometer data to obtain preprocessed optical time domain reflectometer data; inputting the preprocessed optical time domain reflectometer data into an SMC-BiLSTM network model for training to obtain a trained SMC-BiLSTM network model; diagnosing the optical time domain reflectometer data to be identified by using the trained SMC-BiLSTM network model to obtain the type and diagnosis result of the optical fiber fault.

2. The method of claim 1, wherein the SMC-BiLSTM network model-based optical fiber fault diagnosis method is characterized by, The original optical time domain reflectometer data is specifically backscattering signal data collected by an optical time domain reflectometer. 3.The SMC-BiLSTM network model-based optical fiber fault diagnosis method according to claim 1, characterized in that, The architecture of the SMC-BiLSTM network model comprises an input layer, a convolutional neural network layer, a bidirectional long short-term memory network layer and a deep neural network layer. The input layer is configured to receive the preprocessed optical time domain reflectometer data and convert the preprocessed optical time domain reflectometer data into an input tensor. The convolutional neural network layer is configured to perform multi-scale local feature extraction and spatial dimension reduction on the input tensor to generate a time sequence feature map. The bidirectional long short-term memory network layer is configured to perform bidirectional time sequence dependence modeling on the time sequence in the time sequence feature map by a parameter sharing mechanism to generate a feature vector fusing global context. The deep neural network layer is configured to perform high-level feature abstraction and fault classification on the feature vector.

4. The method of claim 3, wherein the SMC-BiLSTM network model is based on a SMC-BiLSTM network model. The convolutional neural network layer comprises a convolutional layer and a pooling layer. The convolutional layer is configured to slide a multi-scale convolutional kernel on the input tensor to extract local spatial features, introduce a non-linear transformation through a ReLU activation function, and generate a time sequence feature map. The pooling layer is configured to down-sample the time sequence feature map to reduce the dimension of the time sequence feature map while retaining important features.

5. The method of claim 3, wherein the SMC-BiLSTM network model-based optical fiber fault diagnosis method is characterized by, The bidirectional long short-term memory network layer comprises: a forward LSTM network and a backward LSTM network configured to capture the forward and backward dependence relationship in the time sequence feature map to generate a feature vector fusing global context.

6. The method of claim 3, wherein the SMC-BiLSTM network model-based optical fiber fault diagnosis method is characterized by, The deep neural network layer comprises a fully connected layer, a dropout layer and an output layer. The fully connected layer is configured to map the feature vector to a high-dimensional space, extract non-linear combination features layer by layer, and introduce a ReLU activation function to enhance the model expression capability. The Dropout layer is configured to prevent the SMC-BiLSTM network model from overfitting by randomly discarding a part of neurons. The output layer is configured to convert the feature vector into a probability distribution vector of the fault type through a Softmax function.

7. The method of claim 3, wherein the SMC-BiLSTM network model-based optical fiber fault diagnosis method is characterized by, The parameters of the bidirectional long short-term memory network layer are consistent when processing all CNN branch outputs, and the total amount of model parameters is reduced through a parameter sharing mechanism.

8. The method of claim 7, wherein the SMC-BiLSTM network model-based optical fiber fault diagnosis method is characterized by, The number of CNN branches is four, and the convolutional kernel scales of the branches are sequentially increased to capture spatial features of different granularities.

9. The method of claim 5, wherein the SMC-BiLSTM network model-based optical fiber fault diagnosis method is characterized by, The process of fusing the global context feature vector is to concatenate the feature dimensions to form the feature vector fusing the global context. 10.A fiber fault diagnosis system based on an SMC-BiLSTM network model, applied to perform the fiber fault diagnosis method based on the SMC-BiLSTM network model according to any one of claims 1-9, characterized in that, The method comprises the following steps: The data acquisition module uses an optical time domain reflectometer to measure the optical fiber and acquires original optical time domain reflectometer data corresponding to different fault types. The multi-scale feature extraction module is configured to preprocess the original optical time domain reflectometer data to obtain preprocessed optical time domain reflectometer data. The model training module is configured to input the preprocessed optical time domain reflectometer data into an SMC-BiLSTM network model for training to obtain a trained SMC-BiLSTM network model. The fault identification module is configured to use the trained SMC-BiLSTM network model to diagnose optical time domain reflectometer data to be identified to obtain a type of optical fiber fault and a diagnosis result.

Images