Method and system for monitoring connection reliability of optical fiber connector based on AI

Abstract

The invention relates to the technical field of optical fiber connector connection monitoring, in particular to an AI-based optical fiber connector connection reliability monitoring method and system, and the method comprises the steps: constructing a multi-mode monitoring data source through synchronously collecting time sequence parameters including optical power, insertion loss, return loss and other physical quantities and a connector end face image, the image is subjected to feature mapping through the convolutional neural network, spatial features of pollution, scratches and alignment deviation are automatically extracted, manual inspection is replaced, and the recognition efficiency and objectivity are improved. Time sequence parameters are input into the modeling performance evolution trend of the long-short-term memory network, and weak dynamic characteristics such as chronic degradation are effectively captured. And through fusing image space features and parameter trend features, strengthening key factor characterization, and outputting a reliability determination result of the connector through a residual error classification network, accurate classification of normal, to-be-maintained and fault states is realized, traceability alarm information can be generated, and intelligent monitoring and early warning of the operation state of the connector are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optical fiber connector connection monitoring, in particular to an AI-based optical fiber connector connection reliability monitoring method and system. BACKGROUND

[0002] As a key component for realizing physical butt joint and transmission of optical signals in optical communication networks, the connection quality of optical fiber connectors is directly related to the transmission stability and performance integrity of the entire network link. In actual operation, due to the long-term open or multiple plugging environment of optical fiber connectors, they are easily affected by factors such as dust pollution, scratch wear, alignment deviation and mechanical looseness, which in turn leads to increased insertion loss, decreased return loss, and even communication interruption or increased bit error rate in severe cases.

[0003] Currently, the monitoring of the connection state of optical fiber connectors mainly relies on periodic manual inspection and physical quantity alarm mechanism based on threshold rules. Among them, manual inspection usually includes visual inspection of fiber end face, connector plugging confirmation and optical path cleaning and maintenance, but has the problems of low operation frequency, poor timeliness and high labor cost. On the other hand, physical quantity monitoring relies on devices such as optical power meter and return loss meter to determine whether the connection state is abnormal by setting a fixed threshold, for example, when the insertion loss exceeds a certain set value, an alarm is triggered, however, there are significant limitations: on the one hand, the threshold setting is highly dependent on experience and is highly sensitive to the environment, which can easily lead to false positives or false negatives in a changing field environment; on the other hand, the threshold mechanism cannot identify the dynamic trend of parameter changes, such as chronic performance degradation caused by end face pollution accumulation or connection looseness, which often causes channel quality to decline before the threshold is triggered. SUMMARY

[0004] To solve the above problems, the present application provides an AI-based optical fiber connector connection reliability monitoring method and system.

[0005] To achieve the above purpose, the technical solution adopted by the present application is:

[0006] An AI-based optical fiber connector connection reliability monitoring method, comprising the following steps:

[0007] S1, collecting multi-modal monitoring data of the optical fiber connector in the communication operation process, the multi-modal monitoring data including time series parameters of optical power, insertion loss and return loss, and connector end face images;

[0008] S2, inputting the end face images into a convolutional neural network model for feature mapping to extract spatial feature vectors of end face pollution, scratches and alignment state; inputting the time series parameters into a long short-term memory network for time series state coding to extract parameter trend feature vectors;

[0009] S3, weighting and fusing the spatial feature vector and the parameter trend feature vector based on an attention mechanism to generate a corresponding connection reliability feature vector;

[0010] S4, inputting the connection reliability feature vector into a residual classification network to output a connection reliability determination result of the fiber connector, the connection reliability determination result including a normal state, a maintenance required state or a fault alarm state;

[0011] S5, generating monitoring alarm information based on the connection reliability determination result.

[0012] Further, the S1 comprises the following steps:

[0013] The original monitoring data of optical power, insertion loss and return loss are collected according to a preset sampling period and a time series data set is constructed;

[0014] The time series data set is normalized and divided into a plurality of subsequences based on a sliding time window to obtain time series parameters;

[0015] The end face of the fiber connector is focused by the camera device to obtain a connector end face image.

[0016] Further, the end face image is input into a convolutional neural network model to extract spatial features of end face pollution, scratches and alignment state, comprising the following steps:

[0017] The obtained connector end face image is subjected to geometric correction, brightness normalization and noise suppression processing to obtain a preprocessed end face image;

[0018] The preprocessed end face image is subjected to multi-convolution feature extraction and batch normalization to obtain an initial texture feature map;

[0019] The initial texture feature map is subjected to dimension reduction and feature compensation in the pooling layer and the residual layer to obtain an enhanced deep feature map;

[0020] According to the enhanced deep feature map, vectorization is performed in the fully connected layer and the pollution area distribution, scratch edge intensity and alignment deviation are quantitatively encoded to obtain a spatial feature vector.

[0021] Further, the time series parameters are input into a long short-term memory network for time series state encoding to extract a parameter trend feature vector, comprising the following steps:

[0022] The time series parameters are divided into equal-length subsequences according to a sliding window, and each subsequence is normalized to obtain a time series input matrix;

[0023] Input the time sequence input matrix into the long short-term memory network, adjust the state of the sequence information through the input gate and the forgetting gate, and obtain a hidden state vector corresponding to a time step;

[0024] Make the hidden state vector perform a nonlinear mapping in the output gate to obtain a time sequence feature sequence;

[0025] Perform feature weighting on the time sequence feature sequence through a weighted aggregation layer, and output a parameter trend feature vector.

[0026] Further, the long short-term memory network is trained by the following steps:

[0027] Collect historical fiber operation samples containing known connection state labels, convert the optical power, insertion loss and return loss of the historical fiber operation samples into training time sequence input matrices, and associate the connection state labels to obtain a supervised training sample set;

[0028] Input the training time sequence input matrix into the long short-term memory network, and output a training parameter trend feature vector;

[0029] Calculate the loss value of the training parameter trend feature vector and the connection state label through the mean square error function, and perform gradient update on the weight parameters of the long short-term memory network through the back propagation algorithm.

[0030] Further, the S3 includes the following steps:

[0031] Concatenate the spatial feature vector and the parameter trend feature vector to form an initial feature set;

[0032] Based on the initial feature set, calculate the attention weight of each feature component and perform weighted summation to obtain a connection reliability feature vector.

[0033] Further, the residual classification network is trained by the following steps:

[0034] Collect historical fiber connector operation samples, and construct a training sample set according to the spatial feature vectors, parameter trend feature vectors and corresponding real connection state labels thereof;

[0035] Input the training sample set into the residual classification network, sequentially extract deep discriminative features through multiple residual convolution units, and output a connection state prediction result;

[0036] Based on the cross entropy loss function between the connection state prediction result and the real connection state label, calculate the training error, and update the model parameters of the residual classification network through the back propagation algorithm until it converges to a preset threshold.

[0037] Further, the cross entropy loss function is as follows:

[0038]

[0039] wherein, L is a loss value; N is the total number of connection reliability state categories; y i is a one-hot encoding of the true state label; p i is the predicted probability of the i-th connection state output by the residual classification network.

[0040] Further, generating monitoring alarm information according to the connection reliability determination result comprises:

[0041] binding the connection reliability determination result with the physical identity of the corresponding connector and the timestamp information to generate an alarm data entry;

[0042] uploading the alarm data entry to a detection platform for alarm action.

[0043] An AI-based optical fiber connector connection reliability monitoring system, applied to any one of the AI-based optical fiber connector connection reliability monitoring methods described above, comprising:

[0044] a data acquisition module for acquiring multi-modal monitoring data of the optical fiber connector during communication operation, the multi-modal monitoring data including time series parameters of optical power, insertion loss and return loss, and connector end face images;

[0045] a feature extraction module for inputting the end face images into a convolutional neural network model for feature mapping, extracting spatial feature vectors of end face contamination, scratches and alignment state; inputting the time series parameters into a long short-term memory network for time series state encoding, extracting parameter trend feature vectors;

[0046] a feature fusion module for weighting and fusing the spatial feature vectors and parameter trend feature vectors based on an attention mechanism to generate corresponding connection reliability feature vectors;

[0047] a detection and determination module for inputting the connection reliability feature vectors into a residual classification network to output a connection reliability determination result of the optical fiber connector, the connection reliability determination result including a normal state, a maintenance state or a fault alarm state;

[0048] an alarm output module for generating monitoring alarm information based on the determination result.

[0049] The application has the beneficial effects that: the application forms a multi-modal monitoring data source containing time sequence data and image information by simultaneously collecting physical parameters such as optical power, insertion loss, return loss and connector end face images, significantly breaks through the limitation of traditional methods based on only single numerical signal monitoring, improves the data comprehensiveness and the accuracy of abnormal capture, further maps the end face image through a convolutional neural network to automatically extract spatial structure features such as pollution, scratches and alignment state, overcomes the problems of strong subjectivity and low efficiency of artificial inspection, and realizes intelligent identification of the end face state; meanwhile, the time sequence parameters are input into a long short-term memory network to model the connector performance change trend, which can effectively mine weak but key time dynamic features such as chronic degradation and periodic jitter, and solve the problem that the static threshold mechanism cannot identify the trend degradation. By weighting and fusing the spatial features and time trend features, the key influencing factors are effectively focused on, and the information dimension imbalance or redundant interference is avoided, on the basis of which, a residual classification network is used to realize deep discrimination of the connection state, and a more stable reliability level label is output, which has higher accuracy and robustness than the traditional threshold judgment method. Finally, traceable monitoring alarm information is generated based on the model judgment result, realizing real-time perception and pre-warning of the connector state change. BRIEF DESCRIPTION OF DRAWINGS

[0050] Fig. 1 is a step flow chart of an AI-based optical fiber connector connection reliability monitoring method in the application.

[0051] Fig. 2 is a training step flow chart of a residual classification network in the application. DETAILED DESCRIPTION

[0052] Referring to Figs. 1-2 The application relates to an AI-based optical fiber connector connection reliability monitoring method, which comprises the following steps:

[0053] S1, collecting multi-modal monitoring data of an optical fiber connector in a communication running process, wherein the multi-modal monitoring data comprises time sequence parameters of optical power, insertion loss and return loss and connector end face images;

[0054] S2, inputting the end face images into a convolutional neural network model for feature mapping to extract spatial feature vectors of end face pollution, scratches and alignment state; inputting the time sequence parameters into a long short-term memory network for time sequence state coding to extract parameter trend feature vectors;

[0055] S3, weighting and fusing the spatial feature vectors and the parameter trend feature vectors based on an attention mechanism to generate corresponding connection reliability feature vectors;

[0056] S4, input the connection reliability feature vector into a residual classification network, output a connection reliability determination result of the optical fiber connector, the connection reliability determination result including a normal state, a maintenance required state or a fault alarm state;

[0057] S5, generate monitoring alarm information based on the connection reliability determination result.

[0058] In some embodiments, the front-end data acquisition device and the edge computing terminal are first deployed to collect the running state data of the fiber connector in real time. Specifically, the front-end data acquisition device collects three types of key data by connecting the fiber link and synchronously collecting: 1) time series physical quantities such as optical power, insertion loss, and return loss, which are completed by a high-speed optical power meter and a reflection loss detector; 2) connector end face images obtained based on an end face microscopic imaging device; and 3) system time stamp and connector number and other identification information, which are used to establish a mapping relationship between the data and the physical object. The above end face images are input into a pre-trained convolutional neural network (CNN) model in the system for feature extraction. The CNN includes a multi-scale convolutional channel structure, which automatically extracts spatial structural features such as surface contamination degree, scratch morphology, and alignment offset angle from the images without relying on manual annotation or rule judgment, significantly improving the recognition accuracy and processing efficiency. At the same time, the time series parameters are input into a long short-term memory network (LSTM) based on a gating mechanism to mine the trend patterns of the parameters over time, such as periodic return loss jitter or slow increase in insertion loss and other implicit degradation phenomena, and then construct a trend feature vector reflecting the dynamic evolution of performance. This processing overcomes the problem that traditional single-point alarm based on fixed threshold cannot identify chronic degradation or precursors of sudden changes. Subsequently, an attention mechanism is introduced for feature fusion processing. Specifically, by constructing an attention weight matrix between the spatial feature vector and the trend feature vector, the contribution value of different feature dimensions in the fusion process is dynamically adjusted, so as to focus on the key factors that have a dominant effect on the change of connection reliability. For example, in the case of severe contamination of the end face and slow increase in loss, the attention mechanism can enhance the fusion expression ability of the contamination feature and the trend change, avoiding model misjudgment due to unbalanced information dimensions. The fused connection reliability feature vector is sent to a deep residual classification network for state discrimination. The network uses a multi-layer residual connection structure to effectively alleviate the gradient vanishing problem in deep network training and improve the nonlinear decision-making ability. The classification output labels include three types of connection states: “normal”, “needs repair”, and “fault alarm”, which have high precision, strong robustness, and sensitive identification ability for boundary states. Finally, according to the classification output result, structured monitoring alarm information is generated and output to the network operation and management platform. This information not only includes the determination result, but also includes the key indicator trend curve, the end face image visual labeling area, and the model confidence score, supporting the operation and maintenance personnel to realize traceable analysis and active operation and maintenance scheduling of the connector state. Compared with the traditional scheme based on static threshold or manual inspection, the data perception dimension integrity, state recognition sensitivity, and early warning proactivity are significantly improved.

[0059] Further, the S1 comprises the following steps:

[0060] Raw monitoring data of optical power, insertion loss and return loss are collected according to a preset sampling period and a time series data set is constructed;

[0061] The time series data set is normalized, and is divided into a plurality of subsequences based on a sliding time window, to obtain time series parameters;

[0062] The end face of the fiber connector is focused by the camera device to obtain a connector end face image.

[0063] In some embodiments, by collecting optical power, insertion loss (IL) and return loss (RL), a multi-dimensional raw monitoring data stream is formed. To ensure the scale consistency of different indicators in subsequent analysis, the system uses a linear normalization function to normalize the numerical value of each type of parameter, maps it to a unified [0, 1] interval, and eliminates the interference of physical magnitude difference on modeling. Subsequently, the system divides the normalized data into time series based on a sliding time window algorithm. By setting the window length as w and the sliding step as s, a set of subsequence covering the entire time series behavior can be constructed by performing a sliding window traversal operation on the original time series. Each subsequence retains data of continuous w time steps, which forms a wxd-dimensional matrix in form, where d is the parameter dimension (such as 3 dimensions, power, IL and RL respectively). This sliding window division strategy not only enhances the local time series modeling capability of the data, but also provides structured input samples for subsequent trend extraction based on LSTM network. At the same time, in terms of end face image acquisition, the system deploys a high-precision automatic focusing microscopic camera device to obtain the visual image of the fiber connector end face. The camera system controls the automatic focusing motor through an edge sharpness optimization function, dynamically evaluates the image sharpness, and automatically triggers image acquisition when the imaging quality reaches the set threshold. The obtained image is subjected to boundary cropping and image enhancement preprocessing (such as adaptive histogram equalization, noise suppression), to ensure that the subsequent convolutional neural network can stably extract spatial features of pollution particles, scratch texture and ferrule alignment.

[0064] Further, the end face image is input into a convolutional neural network model to extract spatial features of end face pollution, scratches and alignment state, including the following steps:

[0065] The obtained connector end face image is subjected to geometric correction, brightness normalization and noise suppression processing to obtain a preprocessed end face image;

[0066] The preprocessed end face image is subjected to multi-convolution feature extraction and batch normalization to obtain an initial texture feature map;

[0067] The initial texture feature map is subjected to dimension reduction and feature compensation in the pooling layer and the residual layer to obtain an enhanced deep feature map;

[0068] According to the enhanced deep feature map, vectorization is performed in a fully connected layer, and a pollution area distribution, a scratch edge intensity, and an alignment deviation are quantitatively encoded to obtain a spatial feature vector.

[0069] In some embodiments, the acquired fiber connector end face image is first input to an image preprocessing module. The module performs image correction through a geometric transformation algorithm (such as affine transformation), ensures that the center area of the end face is aligned with the field of view of the image, and avoids feature extraction errors caused by angle deviation; at the same time, the brightness normalization method (such as histogram equalization or Retinex algorithm) is used to unify the image brightness distribution, and the composite noise suppression is performed by combining the Gaussian filter and the median filter to retain the key texture edge features, and finally the preprocessed end face image with stable structure and clear contrast is output. Subsequently, the preprocessed image is input to the convolutional neural network (CNN) main structure. The main part adopts a feature extraction path stacked with multiple convolution units, each layer containing a convolution layer, a batch normalization (BN) layer and a ReLU activation function, and the convolution kernel size is set to be small (such as 3x3) in the initial stage to capture fine-grained texture features. After being extracted by multiple convolution layers, an initial texture feature map is formed, which contains local pattern expressions such as scratch morphology, pollution patches and fiber edges. In order to solve the problem of feature degradation or gradient disappearance in deep expression of traditional CNN, a residual connection mechanism (ResNet structure) is introduced in this embodiment to construct a jump connection path, so that the deep output and the shallow output are fused by feature compensation. At the same time, the maximum pooling layer is used to realize spatial dimension reduction sampling while keeping the key information, and the robustness of the model under the change of shape scale is improved. Through the above dimension reduction and residual enhancement processing, an enhanced deep feature map with high semantic expression ability is obtained. In the feature vector construction stage, the enhanced deep feature map is input to the fully connected layer for vectorization coding, and special quantization channels are designed according to the task requirements: on the one hand, the spatial distribution density of the pollution area is quantized by a region segmentation method based on pixel clustering (such as SLIC or U-Net auxiliary segmentation); on the other hand, the edge intensity distribution is calculated by extracting the scratch contour line through an edge detection algorithm (such as Sobel + non-maximum suppression); in addition, the alignment deviation value is calculated by the position deviation between the center of the fiber end face and the image boundary. Finally, these indicators are packaged into a multi-dimensional spatial feature vector as the input of the connection state discrimination.

[0070] Further, the inputting the time sequence parameter into a long short-term memory network for time sequence state coding to extract a parameter trend feature vector comprises the following steps:

[0071] Dividing the time sequence parameter into equal-length subsequences according to a sliding window, and performing normalization processing on each subsequence to obtain a time sequence input matrix;

[0072] input the time sequence input matrix into a long short-term memory network, adjust the state of sequence information through an input gate and a forgetting gate, and obtain a hidden state vector corresponding to a time step;

[0073] perform a nonlinear mapping on the hidden state vector in an output gate to obtain a time sequence feature sequence;

[0074] perform feature weighting on the time sequence feature sequence through a weighted aggregation layer, and output a parameter trend feature vector.

[0075] Specifically, first, the original monitoring parameters such as optical power, insertion loss and return loss are formed into time sequences according to fixed time intervals, and the time sequences are divided into equal-length subsequences in a sliding window manner in the preprocessing stage, facilitating the subsequent model to learn the parameter change trend in each time period. Each subsequence is normalized to eliminate the dimensional differences of different physical quantities and make them have a unified feature expression basis. Subsequently, the normalized time sequence subsequences are input into the long short-term memory network (LSTM) structure constructed. The LSTM model has an internal memory mechanism and can dynamically adjust the long-term and short-term information in the time sequence. In this process, the input of each time step not only interacts with the observation value at the current time, but also combines the memory state accumulated at the previous time step to dynamically adjust the current memory output. This mechanism enables the model to identify abnormal patterns or potential trends in the time sequence such as optical power fluctuations and insertion loss mutations. Specifically, at each time step, the model judges the importance of the current input through the input gate, judges whether to retain the previous memory through the forgetting gate, and decides the output mode of the current memory through the output gate. As the entire time sequence slides through the model one by one, the model will output a series of hidden state vectors, each corresponding to the feature representation of each time step after considering the history and current input. To further improve the accuracy of trend modeling, a feature weighting mechanism is introduced after the LSTM output. This mechanism will allocate different weights according to the contribution of each time step hidden state to the overall trend of the sequence, in order to enhance the features of key time periods. Finally, all the hidden state vectors of the time steps are aggregated in a weighted manner to output a global trend feature vector. Through the above method, not only the continuity and potential laws in the process of optical fiber connection state change can be effectively mined, but also the key change segments in the sequence can be dynamically focused on, improving the recognition ability of connector state degradation, failure precursor and other phenomena.

[0076] Further, the long short-term memory network is trained by the following steps:

[0077] Collect historical optical fiber operation samples containing known connection state labels, convert the optical power, insertion loss and return loss of the historical optical fiber operation samples into training time sequence input matrices, and associate the connection state labels to obtain a supervised training sample set;

[0078] inputting the training time sequence input matrix into the long short-term memory network to output a training parameter trend feature vector;

[0079] calculating a loss value of the training parameter trend feature vector and the connection state label through a mean square error function, and performing gradient update on the weight parameters of the long short-term memory network through a back propagation algorithm.

[0080] In some embodiments, first, historical fiber connector operation samples with known connection state labels are collected from an operating optical network. Each sample contains time series records of optical power, insertion loss and return loss, and its corresponding connection state label such as "normal", "needs repair" or "fault" is provided by manual inspection, system log or expert annotation. The above physical quantities are processed by sliding window technology, uniformly divided into equal length sequences, and normalized and standardized preprocessing is performed, thereby constructing a time sequence input matrix for training. Subsequently, the time sequence input matrix is input into the long short-term memory network model constructed. The LSTM structure dynamically controls the state propagation of each time step through its internal input gate, forget gate and output gate, wherein the forget gate is used to determine whether to retain or discard historical information, the input gate controls the degree of writing of new information at the current time step, and the output gate determines the final state output of the current time step, thereby effectively modeling the time variation trend of the connector performance, including key behavior patterns such as chronic degradation, sudden disturbance and periodic fluctuation. The output vector of each sample generated during the forward propagation process of the model is the corresponding "parameter trend feature representation". By setting the loss function as mean square error (MSE), the difference between the trend vector and the actual label is compared to calculate the sample loss. Subsequently, the back propagation algorithm is used in combination with the gradient descent method to iteratively update the internal parameters of the model (including the gating weight and the bias term). Early stopping mechanism and cross-validation strategy are introduced during the training process to avoid overfitting and improve the generalization ability. Compared with the traditional abnormality judgment mechanism based on physical threshold, the embodiment automatically learns the time dynamic evolution features through LSTM, which can identify the connectors in the trend deterioration state before reaching the abnormal threshold, effectively improving the sensitivity and accuracy of fault warning.

[0081] Further, the S3 comprises the following steps:

[0082] concatenating the spatial feature vector and the parameter trend feature vector to form an initial feature set;

[0083] Based on the initial feature set, the connection reliability feature vector is obtained by calculating the attention weight of each feature component and performing weighted summation.

[0084] Specifically, the spatial feature vector extracted in the previous step and the parameter trend feature vector are first spliced by dimension to form an initial feature set with unified representation. This splicing operation not only retains the static structural features such as contamination, scratches and alignment state in the end face image, but also retains the dynamic evolution characteristics of the connector physical performance parameters in the time dimension. However, the directly spliced feature vectors have the hidden dangers of scale difference and uneven contribution between different modalities. Therefore, the embodiment introduces an attention mechanism for feature fusion. Specifically, for each dimension component in the initial feature set, an attention score function is constructed, and the attention weight of each component is calculated combined with the self-attention mechanism. The scoring function can be implemented by a feedforward neural network based on trainable parameters or a scoring method based on dot product. Subsequently, the normalized attention weight is used to weight and sum the component features, and a one-dimensional representation of the connector reliability feature vector is output. This vector is a deep representation of the connector state evaluation, which not only integrates image and time series information, but also significantly suppresses the interference of redundant dimensions and the influence of inefficient features through the attention mechanism.

[0085] Further, the residual classification network is trained by the following steps:

[0086] Collect historical optical fiber connector operation samples, and construct a training sample set according to the spatial feature vector, the parameter trend feature vector and the corresponding real connection state label;

[0087] Input the training sample set into the residual classification network, and sequentially extract deep discriminative features through multiple residual convolution units to output the connection state prediction result;

[0088] Based on the cross-entropy loss function between the connection state prediction result and the real connection state label, the training error is calculated, and the model parameters of the residual classification network are updated through the back propagation algorithm until it converges to a preset threshold.

[0089] In some embodiments, first, a training sample set containing multi-class connection states is constructed, each sample consisting of a spatial feature vector extracted from the connector end face image, a parameter trend feature vector obtained based on a long short-term memory network, and a corresponding true connection state label. The sample feature vector is input into a residual classification network after splicing, which is composed of multiple groups of residual convolution units stacked, each group of residual units containing two convolution layers and an identity mapping branch. The stability of feature gradient in forward propagation and back propagation is maintained through identity mapping, effectively alleviating the gradient vanishing problem in deep network training. In the forward propagation process, the input features are sequentially encoded into high-dimensional discriminative features in each residual unit, and finally the corresponding connection state prediction vector is obtained at the output layer. To train the network, the cross-entropy loss function with Softmax is used to quantify the difference between the predicted value and the true label, and the back propagation algorithm is used to perform gradient update on the weight parameters of each convolution kernel. By setting the early stopping mechanism and the optimal model retention strategy, the overfitting phenomenon is prevented. The training process continues to iterate until the loss function converges to the preset threshold. The residual classification network has good discriminant robustness and can significantly improve the recognition accuracy of fiber connection state, especially in the multi-source heterogeneous feature fusion scene, it shows better performance than traditional classifiers.

[0090] Further, the cross-entropy loss function is as follows:

[0091]

[0092] Wherein, L is the loss value; N is the total number of connection reliability state categories; y i is the one-hot encoding of the true state label; p i is the prediction probability of the i-th connection state output by the residual classification network.

[0093] Specifically, for each training sample, first, its corresponding connection state label is one-hot encoded, so that each state corresponds to a unique activated component. After forward propagation of the residual classification network, the prediction probability distribution of each category is output, representing the confidence of the sample belonging to each connection state category. Then, the cross-entropy loss function is used to calculate the loss value between the prediction probability distribution and the one-hot encoded label. The loss function traverses all categories one by one, and imposes higher gradient penalty on the category position corresponding to the true label, thereby encouraging the model to output higher prediction probability at that position. This training method based on probability distribution distance minimization can effectively improve the model's ability to distinguish between "normal", "needs repair" and "fault alarm" connection states. Combined with the enhanced deep feature expression ability of the residual structure, through the back propagation algorithm, the network parameters are continuously adjusted during the multi-round training process, so that the classification network has higher robustness and accuracy when determining complex state boundaries, thereby effectively meeting the engineering needs of real-time monitoring and reliability warning of fiber connector connection states.

[0094] Further, generating monitoring alarm information according to the connection reliability determination result comprises:

[0095] Binding the connection reliability determination result with the physical identifier and timestamp information of the corresponding connector to generate an alarm data entry;

[0096] According to the alarm data entry, upload to the detection platform for alarm action.

[0097] Specifically, first, the connection reliability determination result output by the residual classification network is used as the core determination basis, and the determination result belongs to the classification label, which usually includes multiple state categories such as "normal state", "needs repair state" and "fault alarm state". In order to ensure the context integrity and device recognizability of the alarm information, the system will extract the unique physical identifier (such as device ID, port number, etc.) of the fiber connector corresponding to the current determination state and the specific timestamp of the determination before generating the alarm data. These auxiliary information and the determination result together form a structured alarm data entry, forming a standard data format, which is convenient for subsequent storage, retrieval and traceability. Then, the generated alarm data entry is uploaded to the centralized detection platform to perform alarm pushing and subsequent processing actions. The uploading process usually uses asynchronous transmission mechanisms such as HTTP API or MQ message queue to ensure low latency and not to block the main determination process. After receiving the alarm data, the detection platform can trigger corresponding events based on the determination state level and historical records, such as generating maintenance tasks, marking as high-priority work orders, or notifying operation and maintenance personnel, etc.

[0098] The application also comprises an AI-based optical fiber connector connection reliability monitoring system, applied to any one of the above-described AI-based optical fiber connector connection reliability monitoring methods, comprising:

[0099] a data acquisition module for acquiring multi-modal monitoring data of the optical fiber connector during communication operation, the multi-modal monitoring data comprising time series parameters of optical power, insertion loss and return loss, and connector end face images;

[0100] a feature extraction module for inputting the end face images into a convolutional neural network model for feature mapping, extracting spatial feature vectors of end face contamination, scratches and alignment state, and inputting the time series parameters into a long short-term memory network for time series state coding, extracting parameter trend feature vectors;

[0101] a feature fusion module for weighting and fusing the spatial feature vectors and parameter trend feature vectors based on an attention mechanism, generating corresponding connection reliability feature vectors;

[0102] a detection and judgment module for inputting the connection reliability feature vectors into a residual classification network, outputting connection reliability judgment results of the optical fiber connector, the connection reliability judgment results comprising normal state, repair required state or fault alarm state;

[0103] an alarm output module for generating monitoring alarm information based on the judgment results.

[0104] The above embodiments only describe the preferred embodiments of the application, and do not limit the scope of the application. Without departing from the design spirit of the application, various modifications and improvements to the technical solutions of the application made by ordinary engineering technicians in the art shall fall within the protection scope determined by the claims of the application.

Claims

1. A method for monitoring the reliability of fiber optic connector connections based on AI, characterized in that, Includes the following steps: S1. Collect multi-mode monitoring data of the fiber optic connector during communication operation. The multi-mode monitoring data includes time-series parameters of optical power, insertion loss and return loss, as well as connector end face images. S2. Input the end face image into a convolutional neural network model for feature mapping to extract spatial feature vectors of end face contamination, scratches, and alignment status; input the time series parameters into a long short-term memory network for temporal state encoding to extract parameter trend feature vectors. S3. Based on the attention mechanism, the spatial feature vector and the parameter trend feature vector are weighted and fused to generate the corresponding connection reliability feature vector; S4. Input the connection reliability feature vector into the residual classification network and output the connection reliability determination result of the fiber optic connector. The connection reliability determination result includes normal state, maintenance required state or fault alarm state. S5. Generate monitoring alarm information based on the connection reliability determination result.

2. The AI-based fiber optic connector connection reliability monitoring method according to claim 1, characterized in that, S1 includes the following steps: The raw monitoring data of optical power, insertion loss and return loss are collected according to the preset sampling period and a time series dataset is constructed. The time series dataset is normalized and divided into several subsequences based on a sliding time window to obtain time series parameters; The end face of the fiber optic connector is focused using a camera device to obtain an image of the connector end face.

3. The method for monitoring the reliability of fiber optic connector connections based on AI according to claim 1, characterized in that, The process of inputting the end-face image into a convolutional neural network model to extract spatial features of end-face contamination, scratches, and alignment status includes the following steps: The acquired connector end face image is subjected to geometric correction, brightness normalization and noise suppression to obtain a preprocessed end face image; Perform multi-convolution feature extraction and batch normalization on the preprocessed end face image to obtain an initial texture feature map; An enhanced deep feature map is obtained by performing dimensionality reduction and feature compensation on the initial texture feature map through pooling layers and residual layers; Based on the enhanced deep feature map, vectorization is performed in the fully connected layer, and the distribution of contaminated areas, scratch edge intensity, and alignment deviation are quantized and encoded to obtain spatial feature vectors.

4. The method for monitoring the reliability of fiber optic connector connections based on AI according to claim 3, characterized in that, The step of inputting the time series parameters into a long short-term memory network for time-series state encoding and extracting parameter trend feature vectors includes the following steps: The time series parameters are divided into equal-length subsequences by a sliding window, and each subsequence is normalized to obtain the time series input matrix. The time-series input matrix is ​​input into a long short-term memory network, and the state of the sequence information is adjusted through the input gate and the forget gate to obtain the hidden state vector of the corresponding time step. The hidden state vector is nonlinearly mapped in the output gate to obtain a temporal feature sequence. The time-series feature sequences are weighted by a weighted aggregation layer, and the output parameter trend feature vector is generated.

5. The method for monitoring the reliability of fiber optic connector connections based on AI according to claim 4, characterized in that, The Long Short-Term Memory network is trained through the following steps: Collect historical fiber operation samples containing known connection status labels, convert the optical power, insertion loss, and return loss of the historical fiber operation samples into a training time series input matrix, and associate them with connection status labels to obtain a supervised training sample set; The training time sequence input matrix is ​​input into the long short-term memory network, and the training parameter trend feature vector is output. The loss value of the training parameter trend feature vector and connection state label is calculated by the mean squared error function, and the gradient update of the weight parameters of the long short-term memory network is performed by the backpropagation algorithm.

6. The method for monitoring the reliability of fiber optic connector connections based on AI according to claim 1, characterized in that, S3 includes the following steps: The spatial feature vector and the parametric trend feature vector are concatenated to form an initial feature set; Based on the initial feature set, the connection reliability feature vector is obtained by calculating the attention weight of each feature component and performing a weighted summation.

7. The method for monitoring the reliability of fiber optic connector connections based on AI according to claim 5, characterized in that, The residual classification network is trained through the following steps: Collect historical fiber optic connector operation samples and construct a training sample set based on their spatial feature vector, parameter trend feature vector, and corresponding real connection status labels; The training sample set is input into the residual classification network, and deep discriminative features are extracted sequentially through multiple residual convolutional units to output the connection state prediction results. The training error is calculated based on the cross-entropy loss function between the predicted connection state and the actual connection state label. The model parameters of the residual classification network are then updated using the backpropagation algorithm until the network converges to a preset threshold.

8. The method for monitoring the reliability of fiber optic connector connections based on AI according to claim 7, characterized in that, The cross-entropy loss function is as follows: Where L is the loss value; N is the total number of connection reliability state categories; y i One-hot encoding for real state labels; p i This represents the predicted probability of the i-th type of connection state output by the residual classification network.

9. The method for monitoring the reliability of fiber optic connector connections based on AI according to claim 1, characterized in that, Based on the connection reliability determination result, the monitoring alarm information generated includes: The connection reliability determination result is bound to the physical identifier and timestamp information of the corresponding connector to generate alarm data entries; The alarm data entries are uploaded to the detection platform to trigger alarm actions.

10. An AI-based fiber optic connector connection reliability monitoring system, applied to the AI-based fiber optic connector connection reliability monitoring method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire multi-mode monitoring data of the fiber optic connector during communication operation. The multi-mode monitoring data includes time-series parameters of optical power, insertion loss and return loss, as well as connector end face images. The feature extraction module is used to input the end face image into a convolutional neural network model for feature mapping, and extract spatial feature vectors of end face contamination, scratches and alignment status; and input the time series parameters into a long short-term memory network for temporal state encoding, and extract parameter trend feature vectors. The feature fusion module is used to perform weighted fusion of the spatial feature vector and the parameter trend feature vector based on an attention mechanism to generate a corresponding connection reliability feature vector. The detection and judgment module is used to input the connection reliability feature vector into the residual classification network and output the connection reliability judgment result of the fiber optic connector. The connection reliability judgment result includes normal state, maintenance required state or fault alarm state. The alarm output module is used to generate monitoring alarm information based on the judgment result.

Images