Optical fiber winding detection method and device

Abstract

The invention discloses an optical fiber winding detection method and device, and the method comprises the steps: obtaining an optical pulse transmission signal transmitted by an optical time domain reflectometer to a to-be-detected optical fiber, and receiving a reflected optical pulse reflection signal; respectively carrying out signal waveform feature extraction on the optical pulse transmission signal and the optical pulse reflection signal to obtain a transmission signal waveform semantic feature map and a reflection signal waveform semantic feature map; processing the transmitted signal waveform semantic feature map and the reflected signal waveform semantic feature map to obtain a transmitted signal waveform semantic saliency feature map and a reflected signal waveform semantic saliency feature map; performing nonlinear correlation interaction processing on the transmitted signal waveform semantic saliency feature map and the reflected signal waveform semantic saliency feature map to obtain a transmitted signal-reflected signal waveform semantic correlation feature map; the classifier outputs an optical fiber winding identification result according to the transmitted signal-reflected signal waveform semantic association feature map; the method can improve the accuracy of optical fiber winding identification.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an optical fiber winding detection method and device. Background Art

[0002] With the rapid development of information technology, fiber-optic communication plays an increasingly important role in modern communication networks due to its advantages such as high bandwidth, low loss, and anti-electromagnetic interference. The construction and maintenance of fiber-optic networks are crucial to ensuring the stability and efficiency of information transmission. However, during the wiring, maintenance, and upgrade of fiber-optic networks, the problem of optical fiber entanglement gradually emerges, which not only increases maintenance costs, but also may affect the transmission quality of signals, resulting in reduced signal transmission efficiency and even communication interruption.

[0003] Patent document CN113554067 B discloses a method and device for detecting fiber winding defects based on ensemble learning, the method comprising the following steps: S1, collecting images of fiber winding; S2, preprocessing the images collected in step S1; S3, performing defect detection on the images preprocessed in step S2 through a digital image processing model; S4, performing defect detection on the images preprocessed in step S2 through an SVM model; S5, voting on the defect detection results of the two models in steps S3 and S4 according to the corresponding weights based on a linear combination of the digital image processing model and the SVM model with set weights, judging whether the voting results for multiple consecutive images are not less than the set threshold score, and if so, judging that there is a fiber winding defect. The method detects fiber winding based on image recognition, but the fiber network has dense wiring and many ports, and the accuracy of the image recognition method is not high. Summary of the invention

[0004] The present invention provides a method and a device for detecting optical fiber winding, which can improve the accuracy of optical fiber winding detection.

[0005] A method for detecting optical fiber winding, comprising:

[0006] Acquire the optical pulse transmission signal sent by the optical time domain reflectometer to the optical fiber to be detected, and receive the reflected optical pulse reflection signal;

[0007] Inputting the optical pulse emission signal and the optical pulse reflection signal into a pre-trained convolutional neural network model, wherein the convolutional neural network model includes a signal waveform feature extractor, a gated screening adaptive attention module, a nonlinear association interaction module, and a classifier;

[0008] The signal waveform feature extractor extracts signal waveform features from the optical pulse transmission signal and the optical pulse reflection signal respectively to obtain a transmission signal waveform semantic feature graph and a reflection signal waveform semantic feature graph;

[0009] The gated screening adaptive attention module processes the semantic feature map of the emission signal waveform and the semantic feature map of the reflection signal waveform to obtain a semantic saliency feature map of the emission signal waveform and a semantic saliency feature map of the reflection signal waveform;

[0010] The nonlinear association interaction module performs nonlinear association interaction processing on the semantically significant feature map of the transmitted signal waveform and the semantically significant feature map of the reflected signal waveform to obtain a semantic association feature map of the transmitted signal-reflected signal waveform;

[0011] The classifier outputs an optical fiber winding recognition result according to the transmitted signal-reflected signal waveform semantic association feature map.

[0012] Furthermore, the gated screening adaptive attention module processes the semantic feature map of the emission signal waveform and the semantic feature map of the reflection signal waveform to obtain the semantic saliency feature map of the emission signal waveform and the semantic saliency feature map of the reflection signal waveform, including:

[0013] The semantic feature map of the transmitted signal waveform and the semantic feature map of the reflected signal waveform are respectively subjected to multi-scale global pooling processing, position weighted sum, bias processing, activation processing, mask processing, normalization operation and weighted optimization to obtain the semantic saliency feature map of the transmitted signal waveform and the semantic saliency feature map of the reflected signal waveform.

[0014] Further, the semantic feature map of the transmitted signal waveform and the semantic feature map of the reflected signal waveform are respectively subjected to multi-scale global pooling processing and position weighted sum, so as to obtain a multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and a multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform;

[0015] The multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform are respectively subjected to bias processing to obtain a bias-adjusted multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and a bias-adjusted multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform;

[0016] The bias-adjusted transmitted signal waveform semantic association feature multi-scale pooling representation vector and the bias-adjusted reflected signal waveform semantic association feature multi-scale pooling representation vector are respectively activated through an activation function to obtain a transmitted signal waveform channel semantic association feature vector and a reflected signal waveform channel semantic association feature vector;

[0017] The transmission signal waveform channel semantic association feature vector and the reflection signal waveform channel semantic association feature vector are respectively input into the gating unit to perform masking and normalization operations on the feature values ​​at each position to obtain a normalized transmission signal waveform semantic weight vector and a normalized reflection signal waveform semantic weight vector;

[0018] The eigenvalues ​​of each position in the normalized transmitted signal waveform semantic weight vector and the normalized reflected signal waveform semantic weight vector are used as weighting coefficients, and weighted optimization is performed on each feature matrix along the channel dimension in the transmitted signal waveform semantic feature map and the reflected signal waveform semantic feature map to obtain the transmitted signal waveform semantic saliency feature map and the reflected signal waveform semantic saliency feature map.

[0019] Furthermore, multi-scale global pooling processing and position weighted sum are performed on the semantic feature map of the transmitted signal waveform and the semantic feature map of the reflected signal waveform, respectively, including:

[0020] Performing global pooling processing based on maximum value, random value and average value on the semantic feature map of the transmitted signal waveform and the semantic feature map of the reflected signal waveform respectively, and obtaining a transmitted signal waveform semantic global maximum pooling feature vector, a transmitted signal waveform semantic global random value pooling feature vector, a transmitted signal waveform semantic global mean pooling feature vector, a reflected signal waveform semantic global maximum pooling feature vector, a reflected signal waveform semantic global random value pooling feature vector and a reflected signal waveform semantic global mean pooling feature vector respectively;

[0021] The transmitted signal waveform semantic global maximum pooling feature vector, the transmitted signal waveform semantic global random value pooling feature vector and the transmitted signal waveform semantic global mean pooling feature vector are weighted according to position to obtain a multi-scale pooling representation vector of the transmitted signal waveform semantic association feature;

[0022] The reflected signal waveform semantic global maximum pooling feature vector, the reflected signal waveform semantic global random value pooling feature vector and the reflected signal waveform semantic global mean pooling feature vector are weighted according to position to obtain a reflected signal waveform semantic association feature multi-scale pooling representation vector.

[0023] Furthermore, after obtaining the multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform, bias processing is performed, including:

[0024] The multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform are respectively multiplied by the weight matrix, and then added with the bias vector to obtain the bias-adjusted multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the bias-adjusted multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform;

[0025] After obtaining the bias-adjusted transmitted signal waveform semantic association feature multi-scale pooling representation vector and the bias-adjusted reflected signal waveform semantic association feature multi-scale pooling representation vector, activation processing is performed through an activation function to obtain the transmitted signal waveform channel semantic association feature vector and the reflected signal waveform channel semantic association feature vector.

[0026] Furthermore, after obtaining the semantically associated feature vector of the transmitted signal waveform channel and the semantically associated feature vector of the reflected signal waveform channel, they are input into a gating unit to perform masking and normalization operations on the feature values ​​at each position, including:

[0027] Based on the gating mechanism, the feature value of each position in the semantic association feature vector of the transmission signal waveform channel is compared with a preset threshold, the feature value greater than or equal to the preset threshold is taken as the original value, and the feature value less than the preset threshold is taken as 0, so as to obtain a masked transmission signal waveform channel semantic association feature vector;

[0028] Compare the feature value of each position in the semantic association feature vector of the reflection signal waveform channel with a preset threshold, take the original value for the feature value greater than or equal to the preset threshold, and take the feature value less than the preset threshold as 0, to obtain a masked reflection signal waveform channel semantic association feature vector;

[0029] Normalizing each eigenvalue in the masked transmission signal waveform channel semantic association feature vector and the masked reflection signal waveform channel semantic association feature vector to obtain a normalized transmission signal waveform semantic weight vector and a normalized reflection signal waveform semantic weight vector, respectively;

[0030] After obtaining the normalized transmitted signal waveform semantic weight vector and the normalized reflected signal waveform semantic weight vector, weighted optimization is performed to obtain a transmitted signal waveform semantic saliency feature map and a reflected signal waveform semantic saliency feature map.

[0031] Furthermore, the nonlinear association interaction module performs nonlinear association interaction processing on the semantically significant feature map of the transmitted signal waveform and the semantically significant feature map of the reflected signal waveform to obtain a semantically significant feature map of the transmitted signal-reflected signal waveform, including:

[0032] Interactively process the semantically saliency feature map of the transmitted signal waveform and the semantically saliency feature map of the reflected signal waveform to obtain a semantically interactive feature map of the transmitted signal and the reflected signal;

[0033] According to the transmission signal-reflection signal semantic interaction feature map and the preset adjustment hyperparameters, a transmission-reflection waveform semantic interaction nonlinear response value is calculated;

[0034] The semantic interaction nonlinear response value of the emission-reflection waveform is reshaped into a feature map to obtain the semantic association feature map of the emission signal-reflection signal waveform.

[0035] Furthermore, the semantically saliency feature map of the transmitted signal waveform and the semantically saliency feature map of the reflected signal waveform are divided position by position to obtain the transmitted signal-reflected signal semantic interaction feature map.

[0036] Further, according to the transmission signal-reflection signal semantic interaction feature map and the preset adjustment hyperparameters, the transmission-reflection waveform semantic interaction nonlinear response value is calculated, including:

[0037] Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by the first preset adjustment hyperparameter to obtain a first transmitted-reflected signal semantic interaction nonlinear factor quadratic term adjustment parameter;

[0038] Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by the second preset adjustment hyperparameter to obtain a first-order adjustment parameter of the nonlinear factor of the first transmitted-reflected signal semantic interaction;

[0039] Adding the first transmission-reflection signal semantic interaction nonlinear factor quadratic term adjustment parameter and the first transmission-reflection signal semantic interaction nonlinear factor linear term adjustment parameter to obtain the first transmission-reflection signal semantic interaction nonlinear factor;

[0040] Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by a third preset adjustment hyperparameter to obtain a second transmission-reflected signal semantic interaction nonlinear factor quadratic term adjustment parameter;

[0041] Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by a fourth preset adjustment hyperparameter to obtain a first-order adjustment parameter of a second transmitted-reflected signal semantic interaction nonlinear factor;

[0042] Adding the second transmission-reflection signal semantic interaction nonlinear factor quadratic term adjustment parameter to the second transmission-reflection signal semantic interaction nonlinear factor linear term adjustment parameter to obtain a second transmission-reflection signal semantic interaction nonlinear factor;

[0043] The first transmission-reflection signal semantic interaction nonlinear factor is divided by the second transmission-reflection signal semantic interaction nonlinear factor to obtain a transmission-reflection waveform semantic interaction nonlinear response value.

[0044] Furthermore, the method further comprises:

[0045] When the optical fiber winding identification result shows that the optical fiber is entangled, a warning message is generated and displayed.

[0046] An optical fiber winding detection device applied to the above method comprises:

[0047] A signal acquisition module is used to acquire the optical pulse transmission signal sent by the optical time domain reflectometer to the optical fiber to be detected, and receive the reflected optical pulse reflection signal;

[0048] A detection module, used for inputting the light pulse emission signal and the light pulse reflection signal into a pre-trained convolutional neural network model, wherein the convolutional neural network model includes a signal waveform feature extractor, a gated screening adaptive attention module, a nonlinear association interaction module and a classifier;

[0049] The signal waveform feature extractor is used to extract signal waveform features from the optical pulse transmission signal and the optical pulse reflection signal respectively, to obtain a transmission signal waveform semantic feature graph and a reflection signal waveform semantic feature graph;

[0050] The gated screening adaptive attention module is used to process the semantic feature map of the emission signal waveform and the semantic feature map of the reflection signal waveform to obtain the semantic saliency feature map of the emission signal waveform and the semantic saliency feature map of the reflection signal waveform;

[0051] The nonlinear association interaction module is used to perform nonlinear association interaction processing on the semantically significant feature map of the transmitted signal waveform and the semantically significant feature map of the reflected signal waveform to obtain a semantic association feature map of the transmitted signal-reflected signal waveform;

[0052] The classifier is used to output the optical fiber winding recognition result according to the transmitted signal-reflected signal waveform semantic association feature map.

[0053] Furthermore, the gated screening adaptive attention module processes the semantic feature map of the emission signal waveform and the semantic feature map of the reflection signal waveform to obtain the semantic saliency feature map of the emission signal waveform and the semantic saliency feature map of the reflection signal waveform, including:

[0054] The semantic feature map of the transmitted signal waveform and the semantic feature map of the reflected signal waveform are respectively subjected to multi-scale global pooling processing, position weighted sum, bias processing, activation processing, mask processing, normalization operation and weighted optimization to obtain the semantic saliency feature map of the transmitted signal waveform and the semantic saliency feature map of the reflected signal waveform.

[0055] Furthermore, multi-scale global pooling processing and position weighted sum are performed on the semantic feature map of the transmitted signal waveform and the semantic feature map of the reflected signal waveform, respectively, including:

[0056] Performing maximum-based global pooling processing, random-value-based global pooling processing, and average-based global pooling processing on the semantic feature graph of the transmitted signal waveform and the semantic feature graph of the reflected signal waveform, respectively, to obtain a transmitted signal waveform semantic global maximum pooling feature vector, a transmitted signal waveform semantic global random value pooling feature vector, a transmitted signal waveform semantic global average pooling feature vector, a reflected signal waveform semantic global maximum pooling feature vector, a reflected signal waveform semantic global random value pooling feature vector, and a reflected signal waveform semantic global average pooling feature vector;

[0057] The transmitted signal waveform semantic global maximum pooling feature vector, the transmitted signal waveform semantic global random value pooling feature vector and the transmitted signal waveform semantic global mean pooling feature vector are weighted according to position to obtain a multi-scale pooling representation vector of the transmitted signal waveform semantic association feature;

[0058] The reflected signal waveform semantic global maximum pooling feature vector, the reflected signal waveform semantic global random value pooling feature vector and the reflected signal waveform semantic global mean pooling feature vector are weighted according to position to obtain a reflected signal waveform semantic association feature multi-scale pooling representation vector.

[0059] Furthermore, after obtaining the multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform, bias processing is performed, including:

[0060] The multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform are respectively multiplied by the weight matrix, and then added with the bias vector to obtain the bias-adjusted multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the bias-adjusted multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform;

[0061] After obtaining the bias-adjusted transmitted signal waveform semantic association feature multi-scale pooling representation vector and the bias-adjusted reflected signal waveform semantic association feature multi-scale pooling representation vector, activation processing is performed through an activation function to obtain the transmitted signal waveform channel semantic association feature vector and the reflected signal waveform channel semantic association feature vector.

[0062] Furthermore, after obtaining the semantically associated feature vector of the transmitted signal waveform channel and the semantically associated feature vector of the reflected signal waveform channel, they are input into a gating unit to perform masking and normalization operations on the feature values ​​at each position, including:

[0063] Based on the gating mechanism, the feature value of each position in the semantic association feature vector of the transmission signal waveform channel is compared with a preset threshold, the feature value greater than or equal to the preset threshold is taken as the original value, and the feature value less than the preset threshold is taken as 0, so as to obtain a masked transmission signal waveform channel semantic association feature vector;

[0064] Compare the feature value of each position in the semantic association feature vector of the reflection signal waveform channel with a preset threshold, take the original value for the feature value greater than or equal to the preset threshold, and take the feature value less than the preset threshold as 0, to obtain a masked reflection signal waveform channel semantic association feature vector;

[0065] Normalizing each eigenvalue in the masked transmission signal waveform channel semantic association feature vector and the masked reflection signal waveform channel semantic association feature vector to obtain a normalized transmission signal waveform semantic weight vector and a normalized reflection signal waveform semantic weight vector, respectively;

[0066] After obtaining the normalized transmitted signal waveform semantic weight vector and the normalized reflected signal waveform semantic weight vector, weighted optimization is performed to obtain a transmitted signal waveform semantic saliency feature map and a reflected signal waveform semantic saliency feature map.

[0067] Furthermore, the nonlinear association interaction module performs nonlinear association interaction processing on the semantically significant feature map of the transmitted signal waveform and the semantically significant feature map of the reflected signal waveform to obtain a semantically significant feature map of the transmitted signal-reflected signal waveform, including:

[0068] Interactively process the semantically saliency feature map of the transmitted signal waveform and the semantically saliency feature map of the reflected signal waveform to obtain a semantically interactive feature map of the transmitted signal and the reflected signal;

[0069] According to the transmission signal-reflection signal semantic interaction feature map and the preset adjustment hyperparameters, a transmission-reflection waveform semantic interaction nonlinear response value is calculated;

[0070] The semantic interaction nonlinear response value of the emission-reflection waveform is reshaped into a feature map to obtain the semantic association feature map of the emission signal-reflection signal waveform.

[0071] Furthermore, the semantically saliency feature map of the transmitted signal waveform and the semantically saliency feature map of the reflected signal waveform are divided position by position to obtain the transmitted signal-reflected signal semantic interaction feature map.

[0072] Further, according to the transmission signal-reflection signal semantic interaction feature map and the preset adjustment hyperparameters, the transmission-reflection waveform semantic interaction nonlinear response value is calculated, including:

[0073] Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by the first preset adjustment hyperparameter to obtain a first transmitted-reflected signal semantic interaction nonlinear factor quadratic term adjustment parameter;

[0074] Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by the second preset adjustment hyperparameter to obtain a first-order adjustment parameter of the nonlinear factor of the first transmitted-reflected signal semantic interaction;

[0075] Adding the first transmission-reflection signal semantic interaction nonlinear factor quadratic term adjustment parameter and the first transmission-reflection signal semantic interaction nonlinear factor linear term adjustment parameter to obtain the first transmission-reflection signal semantic interaction nonlinear factor;

[0076] Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by a third preset adjustment hyperparameter to obtain a second transmission-reflected signal semantic interaction nonlinear factor quadratic term adjustment parameter;

[0077] Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by a fourth preset adjustment hyperparameter to obtain a first-order adjustment parameter of a second transmitted-reflected signal semantic interaction nonlinear factor;

[0078] Adding the second transmission-reflection signal semantic interaction nonlinear factor quadratic term adjustment parameter to the second transmission-reflection signal semantic interaction nonlinear factor linear term adjustment parameter to obtain a second transmission-reflection signal semantic interaction nonlinear factor;

[0079] The first transmission-reflection signal semantic interaction nonlinear factor is divided by the second transmission-reflection signal semantic interaction nonlinear factor to obtain a transmission-reflection waveform semantic interaction nonlinear response value.

[0080] Furthermore, the device also includes an early warning module, which is used to generate and display early warning information when the optical fiber winding identification result shows that the optical fiber is entangled.

[0081] The optical fiber winding detection method and device provided by the present invention have at least the following beneficial effects:

[0082] (1) Sending optical pulses through an optical time domain reflectometer, and collecting optical pulse emission signals and optical pulse reflection signals, and then introducing artificial intelligence-based signal processing and analysis algorithms to analyze and nonlinearly correlate the optical pulse emission signals and optical pulse reflection signals, so as to identify the implicit pattern and correlation feature information between the optical pulse emission signals and the optical pulse reflection signals, and thus determine whether the fiber entanglement problem occurs. In this way, the entanglement problem in the optical fiber link can be quickly identified and located, and the automatic identification of the fiber entanglement problem can be realized. When the fiber entanglement is detected, an entanglement warning prompt is generated, so that corresponding measures can be taken to effectively prevent the fiber entanglement;

[0083] (2) Through the processing of the gated filtering adaptive attention module, the features that are most critical for identifying fiber entanglement problems can be enhanced, while those less important information can be suppressed. In this way, the model can more clearly identify abnormal patterns in the signal. In other words, the gated filtering adaptive attention module can automatically adjust its focus according to the specific content of the signal. This dynamic selection capability enables the model to adapt to the various entanglement situations that may exist in the optical fiber link, which enables the model to better handle the personalized characteristics of different optical fiber links, so that the model can more accurately identify signs of optical fiber entanglement, thereby improving the accuracy of entanglement detection;

[0084] (3) Through the nonlinear correlation interaction module, the feature maps of the transmitted signal and the reflected signal can be fused. This fusion helps to capture the complex nonlinear relationship between the semantics of the two signals, thereby more comprehensively understanding the state of the optical fiber link and more accurately identifying the entanglement problem in the optical fiber link. In addition, the nonlinear correlation interaction of the semantic features of the transmitted signal waveform and the semantic features of the reflected signal waveform helps to improve the adaptability of the model to different types of optical fiber links, making the detection results more robust.

[0085] (4) During the model training process, by training the classifier through the relevant associated loss function value, it is possible to model the structural detail dependency relative to the macroscopic complexity in the classification scenario, thereby promoting the understanding of the macroscopic complex feature expression of the sample based on the structural details and improving the accuracy of the classification results obtained by the classifier-based fiber winding detector. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1The present invention provides a flowchart of an embodiment of the optical fiber winding detection method.

[0087] Figure 2 A flowchart of an embodiment of performing saliency processing on a waveform semantic feature graph in the optical fiber winding detection method provided by the present invention.

[0088] Figure 3 The present invention provides a flowchart of an embodiment of a nonlinear correlation interaction processing method in an optical fiber winding detection method provided by the present invention.

[0089] Figure 4 A schematic structural diagram of an embodiment of an optical fiber winding detection device provided by the present invention.

[0090] Figure 5 A schematic diagram of the structure of an embodiment of a convolutional neural network in the optical fiber winding detection method provided by the present invention. DETAILED DESCRIPTION

[0091] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0092] refer to Figure 1 In some embodiments, a method for detecting optical fiber winding is provided, comprising:

[0093] S1, obtaining an optical pulse transmission signal sent by an optical time domain reflectometer to the optical fiber to be detected, and receiving a reflected optical pulse reflection signal;

[0094] S2, inputting the light pulse emission signal and the light pulse reflection signal into a pre-trained convolutional neural network model, wherein the convolutional neural network model includes a signal waveform feature extractor, a gated screening adaptive attention module, a nonlinear association interaction module, and a classifier;

[0095] S3, the signal waveform feature extractor extracts signal waveform features from the optical pulse emission signal and the optical pulse reflection signal respectively to obtain a semantic feature graph of the emission signal waveform and a semantic feature graph of the reflection signal waveform;

[0096] S4, the gated screening adaptive attention module processes the semantic feature map of the emission signal waveform and the semantic feature map of the reflection signal waveform to obtain a semantic saliency feature map of the emission signal waveform and a semantic saliency feature map of the reflection signal waveform;

[0097] S5, the nonlinear association interaction module performs nonlinear association interaction processing on the semantic saliency feature map of the transmitted signal waveform and the semantic saliency feature map of the reflected signal waveform to obtain a semantic association feature map of the transmitted signal-reflected signal waveform;

[0098] S6. The classifier outputs a fiber winding recognition result according to the transmitted signal-reflected signal waveform semantic association feature map.

[0099] Specifically, in step S1, first, an optical pulse transmission signal sent by an optical time domain reflectometer (OTDR) is obtained, and a reflected optical pulse reflection signal is received. It should be understood that in optical fiber communication networks, OTDR (optical time domain reflectometer) is a commonly used device for detecting events in optical fiber links, such as breakpoints, connection points, or bending points. The optical time domain reflectometer works by sending an optical pulse and measuring the reflected optical signal. These signals contain rich information about the optical fiber link, but the original signals are often complex and difficult to analyze directly.

[0100] Furthermore, in step S2, the optical pulse emission signal and the optical pulse reflection signal are input into a pre-trained convolutional neural network model, and the convolutional neural network model includes a signal waveform feature extractor, a gated screening adaptive attention module, a nonlinear association interaction module, and a classifier. Specifically, in order to effectively capture features and perform association analysis on the optical pulse emission signal and the optical pulse reflection signal to determine whether optical fiber winding occurs, in an embodiment, the optical pulse emission signal and the optical pulse reflection signal are processed by an improved convolutional neural network model to respectively capture waveform semantic feature information in the optical pulse emission signal and the optical pulse reflection signal, so as to more accurately reflect the actual situation in the optical fiber link, including the physical properties of the optical fiber and potential winding problems, thereby realizing rapid identification and positioning of the optical fiber winding problem.

[0101] Furthermore, in step S3, the optical pulse emission signal and the optical pulse reflection signal are input into the signal waveform feature extractor of the convolutional neural network model for preliminary feature extraction, and the emission signal waveform semantic feature map and the reflection signal waveform semantic feature map are obtained. However, considering that the emission signal waveform semantic feature map and the reflection signal waveform semantic feature map respectively contain waveform semantic features related to the optical pulse emission signal and the optical pulse reflection signal, but when expressing the waveform semantic features of the two, there are not only feature information related to the optical fiber winding problem identification task, but also some other semantic features that contribute less to the task. Based on this, in order to enable the convolutional neural network model to identify the signal features most relevant to the optical fiber winding problem, thereby further improving the accuracy of optical fiber winding identification and positioning, in this embodiment, the emission signal waveform semantic feature map and the reflection signal waveform semantic feature map are further subjected to significant feature extraction.

[0102] Further, refer to Figure 2In step S4, the gated screening adaptive attention module processes the semantic feature map of the emission signal waveform and the semantic feature map of the reflection signal waveform to obtain the semantic saliency feature map of the emission signal waveform and the semantic saliency feature map of the reflection signal waveform, including:

[0103] S41, performing multi-scale global pooling processing and position weighted sum on the semantic feature graph of the transmitted signal waveform and the semantic feature graph of the reflected signal waveform, respectively, to obtain a multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and a multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform;

[0104] S42, performing bias processing on the multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform, respectively, to obtain a bias-adjusted multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and a bias-adjusted multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform;

[0105] S43, respectively, activating the bias-adjusted transmitted signal waveform semantic association feature multi-scale pooling representation vector and the bias-adjusted reflected signal waveform semantic association feature multi-scale pooling representation vector through an activation function to obtain a transmitted signal waveform channel semantic association feature vector and a reflected signal waveform channel semantic association feature vector;

[0106] S44, respectively inputting the semantic association feature vector of the transmission signal waveform channel and the semantic association feature vector of the reflection signal waveform channel into the gating unit to perform masking and normalization operations on the feature values ​​at each position to obtain a normalized transmission signal waveform semantic weight vector and a normalized reflection signal waveform semantic weight vector;

[0107] S45. Using the eigenvalues ​​of each position in the normalized transmitted signal waveform semantic weight vector and the normalized reflected signal waveform semantic weight vector as weighting coefficients, respectively, weighted optimization is performed on each feature matrix along the channel dimension in the transmitted signal waveform semantic feature map and the reflected signal waveform semantic feature map, to obtain the transmitted signal waveform semantic saliency feature map and the reflected signal waveform semantic saliency feature map.

[0108] Specifically, in optical fiber communication, the optical pulse emission signal and the optical pulse reflection signal may contain different information. The gated screening adaptive attention module can process these two signals separately and extract their respective key features so as to further combine these features for comprehensive analysis. Through the processing of the gated screening adaptive attention module, the features that are most critical for identifying optical fiber winding problems can be enhanced, while suppressing less important information. In this way, the convolutional neural network model can more clearly identify abnormal patterns in the signal. In other words, the gated screening adaptive attention module can automatically adjust its focus according to the specific content of the signal. This dynamic selection capability enables the convolutional neural network model to adapt to the various winding situations that may exist in the optical fiber link, which enables the convolutional neural network model to better process the personalized features of different optical fiber links, so that the convolutional neural network model can more accurately identify signs of optical fiber winding, thereby improving the accuracy of winding detection.

[0109] Further, in step S41, multi-scale global pooling processing and position weighted sum are performed on the semantic feature map of the transmitted signal waveform and the semantic feature map of the reflected signal waveform respectively to obtain a multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and a multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform, including:

[0110] S411, respectively performing global pooling processing based on maximum value, global pooling processing based on random value, and global pooling processing based on average value on the semantic feature graph of the transmitted signal waveform and the semantic feature graph of the reflected signal waveform, to obtain a transmitted signal waveform semantic global maximum pooling feature vector, a transmitted signal waveform semantic global random value pooling feature vector, a transmitted signal waveform semantic global mean pooling feature vector, a reflected signal waveform semantic global maximum pooling feature vector, a reflected signal waveform semantic global random value pooling feature vector, and a reflected signal waveform semantic global mean pooling feature vector, respectively;

[0111] S412, performing weighted sum according to position between the global maximum pooling feature vector of the transmitted signal waveform semantics, the global random value pooling feature vector of the transmitted signal waveform semantics, and the global mean pooling feature vector of the transmitted signal waveform semantics, to obtain a multi-scale pooling representation vector of the transmitted signal waveform semantic association feature;

[0112] S413, performing weighted sum according to position between the reflected signal waveform semantic global maximum pooling feature vector, the reflected signal waveform semantic global random value pooling feature vector and the reflected signal waveform semantic global mean pooling feature vector to obtain a reflected signal waveform semantic association feature multi-scale pooling representation vector.

[0113] Further, in step S42, the multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform are respectively subjected to offset processing, including:

[0114] The multi-scale pooling representation vector of the semantic association features of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association features of the reflected signal waveform are multiplied by the weight matrix respectively, and then added with the bias vector to obtain the bias-adjusted multi-scale pooling representation vector of the semantic association features of the transmitted signal waveform and the bias-adjusted multi-scale pooling representation vector of the semantic association features of the reflected signal waveform respectively.

[0115] Specifically, in step S43, the bias-adjusted transmitted signal waveform semantic association feature multi-scale pooling representation vector and the bias-adjusted reflected signal waveform semantic association feature multi-scale pooling representation vector are activated by a sigmoid activation function to obtain a transmitted signal waveform channel semantic association feature vector and a reflected signal waveform channel semantic association feature vector.

[0116] Further, in step S44, the semantically associated feature vector of the transmitted signal waveform channel and the semantically associated feature vector of the reflected signal waveform channel are respectively input into the gating unit to perform masking and normalization operations on the feature values ​​at each position, including:

[0117] S441, based on the gating mechanism, compare the feature value of each position in the semantic association feature vector of the transmission signal waveform channel with a preset threshold, take the feature value greater than or equal to the preset threshold as the original value, and take the feature value less than the preset threshold as 0, to obtain a masked transmission signal waveform channel semantic association feature vector;

[0118] S442, comparing the feature value of each position in the semantic association feature vector of the reflected signal waveform channel with a preset threshold, taking the feature value greater than or equal to the preset threshold as the original value, and taking the feature value less than the preset threshold as 0, to obtain a masked reflected signal waveform channel semantic association feature vector;

[0119] S443. Normalize each eigenvalue in the masked transmitted signal waveform channel semantic association feature vector and the masked reflected signal waveform channel semantic association feature vector to obtain a normalized transmitted signal waveform semantic weight vector and a normalized reflected signal waveform semantic weight vector, respectively.

[0120] In some embodiments, the adaptive attention saliency enhancement formula for the signal by the gated screening adaptive attention module is:

[0121] V c=a·AvgPool(F)+β·MaxPool(F)+γ·StoPool(F); (1)

[0122] A=σ(W a *V c +B a ); (2)

[0123] A m =mask(A); (3)

[0124]

[0125] F′=A′⊙F; (5)

[0126] Wherein, F is the feature map to be processed, which can be the semantic feature map of the transmitted signal waveform and the semantic feature map of the reflected signal waveform. AvgPool(·), MaxPool(·) and StoPool(·) are global mean pooling, maximum value pooling and random value pooling for each feature matrix along the channel dimension of the feature map, respectively. a, β and γ are all preset trainable weight values. V c is the multi-scale pooling representation vector of the semantic features of the transmitted signal (reflected signal) waveform, W a and B a are the weight matrix and bias vector respectively, σ is the sigmoid activation function, A is the semantic association feature vector of the transmitted signal (reflected signal) waveform channel, mask(·) represents the masking process, A m is the masked transmission signal (reflection signal) waveform channel semantic association feature vector, is the eigenvalue of the ith position in the semantic association feature vector of the masked transmitted signal (reflected signal) waveform channel, n represents the number of eigenvalues ​​in the semantic association feature vector of the masked transmitted signal (reflected signal) waveform channel, A i ′ is the eigenvalue of the i-th position in the normalized semantic weight vector of the transmitted signal (reflected signal) waveform, A′ is the normalized semantic weight vector of the transmitted signal (reflected signal) waveform, ⊙ represents the weighted multiplication along the channel dimension, and F′ is the semantic saliency feature map of the transmitted signal (reflected signal) waveform.

[0127] Among them, the weight matrix W a and the bias vector B a It is preset at the beginning of the convolutional neural network model and adjusted during the training process.

[0128] Since both the optical pulse emission signal and the optical pulse reflection signal contain rich information and events about the optical fiber link, such as breakpoints, connection points or bending points, this information helps to identify the fiber winding problem. In particular, considering that there are implicit association patterns and correlation features between the optical pulse emission signal and the optical pulse reflection signal, this association pattern can further improve the accuracy of identifying the fiber winding problem. Based on this, in order to more comprehensively understand the state of the optical fiber link, in an embodiment, the semantically significant feature map of the emission signal waveform and the semantically significant feature map of the reflection signal waveform are further input into a nonlinear association interaction module to obtain a semantically significant feature map of the emission signal-reflection signal waveform. Through the nonlinear association interaction module, the feature maps of the emission signal and the reflection signal can be fused. This fusion helps to capture the complex nonlinear relationship between the semantics of the two signals, thereby more comprehensively understanding the state of the optical fiber link, which is conducive to more accurately identifying the winding problem in the optical fiber link. In addition, by performing nonlinear association interaction on the semantic features of the waveform of the optical pulse emission signal and the semantic features of the waveform of the optical pulse reflection signal, it helps to improve the adaptability of the model to different types of optical fiber links, making the detection results more robust.

[0129] Further, refer to Figure 3 In step S5, the nonlinear association interaction module performs nonlinear association interaction processing on the semantically significant feature map of the transmitted signal waveform and the semantically significant feature map of the reflected signal waveform to obtain a semantically significant feature map of the transmitted signal-reflected signal waveform, including:

[0130] S51, interactively processing the semantically saliency feature graph of the transmitted signal waveform and the semantically saliency feature graph of the reflected signal waveform to obtain a transmitted signal-reflected signal semantic interaction feature graph;

[0131] S52, calculating and obtaining a nonlinear response value of semantic interaction between emission and reflection waveforms according to the emission signal-reflection signal semantic interaction feature map and preset adjustment hyperparameters;

[0132] S53, reshaping the feature map of the semantic interaction nonlinear response value of the emission-reflection waveform to obtain the semantic association feature map of the emission signal-reflection signal waveform.

[0133] Specifically, in step S51, the semantically saliency feature map of the transmitted signal waveform and the semantically saliency feature map of the reflected signal waveform are divided position by position to obtain the transmitted signal-reflected signal semantic interaction feature map.

[0134] Further, in step S52, according to the transmission signal-reflection signal semantic interaction feature map and the preset adjustment hyperparameters, the transmission-reflection waveform semantic interaction nonlinear response value is calculated, including:

[0135] S521, multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by a first preset adjustment hyperparameter to obtain a first transmitted-reflected signal semantic interaction nonlinear factor quadratic term adjustment parameter;

[0136] S522, multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by a second preset adjustment hyperparameter to obtain a first-order adjustment parameter of the nonlinear factor of the first transmitted-reflected signal semantic interaction;

[0137] S523: Add the first transmission-reflection signal semantic interaction nonlinear factor quadratic term adjustment parameter and the first transmission-reflection signal semantic interaction nonlinear factor linear term adjustment parameter to obtain a first transmission-reflection signal semantic interaction nonlinear factor;

[0138] S524, multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by a third preset adjustment hyperparameter to obtain a second transmission-reflected signal semantic interaction nonlinear factor quadratic term adjustment parameter;

[0139] S525, multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by a fourth preset adjustment hyperparameter to obtain a first-order adjustment parameter of a second transmitted-reflected signal semantic interaction nonlinear factor;

[0140] S526, adding the second transmission-reflection signal semantic interaction nonlinear factor quadratic term adjustment parameter to the second transmission-reflection signal semantic interaction nonlinear factor linear term adjustment parameter to obtain a second transmission-reflection signal semantic interaction nonlinear factor;

[0141] S527 . Divide the first transmission-reflection signal semantic interaction nonlinear factor by the second transmission-reflection signal semantic interaction nonlinear factor to obtain a transmission-reflection waveform semantic interaction nonlinear response value.

[0142] Specifically, in the above steps, the semantically significant feature map of the light pulse emission signal waveform and the semantically significant feature map of the light pulse reflection signal waveform are input into the nonlinear association interaction module and processed by the following nonlinear association interaction formula to obtain the semantically significant feature map of the light pulse emission signal-reflection signal waveform

[0143] Wherein, the nonlinear response optimization formula is:

[0144]

[0145]

[0146] in, is the feature value of the (i, j, k)th position in the semantic saliency feature map of the transmitted signal waveform, is the feature value of the (i, j, k)th position in the semantically significant feature map of the reflected signal waveform, is the feature value of the (i, j, k)th position in the semantic interaction feature map of the emission-reflection signal waveform, α, β, γ and ε are adjustment hyperparameters, and f (i,j,k) is the feature value of the (i, j, k)th position in the semantic association feature map of the transmitted signal-reflected signal waveform.

[0147] Then, the transmission signal-reflection signal waveform semantic association feature map is input into the classifier to obtain a detection result, and the detection result is used to indicate whether entanglement occurs. In other words, the association pattern and correlation feature between the waveform semantic feature of the optical pulse transmission signal and the waveform semantic feature of the optical pulse reflection signal are used to determine whether the optical fiber entanglement problem occurs. In this way, the entanglement problem in the optical fiber link can be quickly identified and located, and the automatic identification of the problem of optical fiber entanglement is realized. In some embodiments, the method further includes:

[0148] When the optical fiber winding identification result shows that the optical fiber is entangled, a warning message is generated and displayed.

[0149] Furthermore, in step S2, the convolutional neural network model is pre-trained, including:

[0150] The invention relates to a method for obtaining a sample optical pulse emission signal sent by an optical time domain reflectometer, receiving a sample optical pulse reflection signal reflected back, and determining whether a true value of optical fiber winding is generated; performing signal waveform feature extraction on the sample optical pulse emission signal and the sample optical pulse reflection signal respectively to obtain a sample emission signal waveform semantic feature map and a sample reflection signal waveform semantic feature map; inputting the sample emission signal waveform semantic feature map and the sample reflection signal waveform semantic feature map into the gated screening adaptive attention module to obtain a sample emission signal waveform semantic saliency feature map and a sample reflection signal waveform semantic saliency feature map; performing gated screening adaptive attention module on the sample emission signal waveform semantic saliency feature map and the sample reflection signal waveform semantic saliency feature map; performing gated screening adaptive attention module on the sample emission signal waveform semantic saliency feature map and the sample reflection signal waveform semantic saliency feature map. The shape semantic saliency feature map is processed nonlinearly with interaction to obtain a sample emission signal-reflection signal waveform semantic association feature map; the sample emission signal-reflection signal waveform semantic association feature map is calculated by the classifier to obtain a classification loss function value; based on the sample emission signal-reflection signal waveform semantic association feature map, the sample emission signal-reflection signal waveform semantic association loss function value is calculated; the signal waveform feature extractor, the gated screening adaptive attention module, the nonlinear association interaction module and the classifier are trained with the weighted sum of the classification loss function value and the sample emission signal-reflection signal waveform semantic association loss function value as the final loss function value.

[0151] In some embodiments, taking into account the misalignment of the source image semantic distribution of the sample light pulse emission signal and the sample light pulse reflection signal, and the gated screening adaptive attention weighted enhancement of the screening adaptive attention for the spatial distribution of local image semantic features by the gated screening adaptive attention module, the sample emission signal waveform semantic saliency feature map and the sample reflection signal waveform semantic saliency feature map will also have image semantic feature mismatch, so that the sample emission signal-reflection signal waveform semantic association feature map obtained by the nonlinear association interaction module will also have image semantic distribution knowledge complexity relative to the source image semantic distribution, so that there is a relatively obvious problem of insufficient classification understanding of complex feature expression, which affects the accuracy of the classification result obtained by the classifier of the sample emission signal-reflection signal waveform semantic association feature map.

[0152] Based on this, the present embodiment further introduces a loss function for promoting classification understanding based on the complex feature expression of the sample emission signal-reflection signal waveform semantic association feature map, that is, based on the sample emission signal-reflection signal waveform semantic association feature map, the sample emission signal-reflection signal waveform semantic association loss function value is calculated. Wherein, based on the sample emission signal-reflection signal waveform semantic association feature map, the sample emission signal-reflection signal waveform semantic association loss function value is calculated, including: first calculating the mean and variance between the eigenvalues ​​of any two positions of the sample emission signal-reflection signal waveform semantic association feature vector obtained after the sample emission signal-reflection signal waveform semantic association feature map is expanded to obtain a mean weight matrix and a variance weight matrix, and then performing query-type multiplication of the sample emission signal-reflection signal waveform semantic association feature vector with the mean weight matrix and the variance weight matrix respectively, and then calculating the Frobenius norm of the autocorrelation matrix of the two eigenvectors obtained, and obtaining the sample emission signal-reflection signal waveform semantic association loss function value by multiplying the Frobenius norm of the autocorrelation matrix with the weight as a hyperparameter.

[0153] For example, represented as:

[0154]

[0155] Where V represents the semantic association feature vector of the sample transmission signal-reflection signal waveform, M μ is the mean weight matrix, M σ is the variance weight matrix, T represents the transpose of the vector, represents the matrix multiplication operation, α is the weight as a hyperparameter, ||·|| F represents the Frobenius norm of the matrix, Loss represents the semantic association loss function value of the sample transmission signal-reflection signal waveform, and as described above:

[0156]

[0157] Among them, v i represents the i-th eigenvalue of the semantic association feature vector of the sample emission signal-reflection signal waveform, v j represents the jth eigenvalue of the semantic association feature vector of the sample transmission signal-reflection signal waveform, M μ (i,j) represents the eigenvalue of the (i,j)th position of the mean weight matrix, M σ (i,j) represents the eigenvalue of the (i,j)th position of the variance weight matrix.

[0158] That is, the sample emission signal-reflection signal waveform semantic association loss function value determines the overall distributed structural bottleneck of the sample emission signal-reflection signal waveform semantic association feature map through the multi-dimensional interaction of the detail group of the sample emission signal-reflection signal waveform semantic association feature map, and simulates the correlation between the structural details of the sample emission signal-reflection signal waveform semantic association feature map and the macro-structured feature representation behavior through the low-rank structured reasoning of the sample emission signal-reflection signal waveform semantic association feature map. In this way, by training the classifier with the sample emission signal-reflection signal waveform semantic association loss function value, it is possible to model the structural detail dependency relative to the macro-complexity in the classification scenario, thereby promoting the understanding of the macro-complex feature expression based on the structural details of the sample emission signal-reflection signal waveform semantic association feature map, and improving the accuracy of the classification results of the classifier. In this way, the winding problem in the optical fiber link can be more accurately identified and located, and the automatic identification of the problem of optical fiber winding is realized, so as to take corresponding measures to effectively prevent the optical fibers from winding with each other.

[0159] Further, refer to Figure 4 and Figure 5 In some embodiments, a fiber winding detection device applied to the above method is provided, comprising:

[0160] The signal acquisition module 201 is used to acquire the optical pulse transmission signal sent by the optical time domain reflectometer to the optical fiber to be detected, and receive the reflected optical pulse reflection signal;

[0161] A detection module 202, used for inputting the optical pulse emission signal and the optical pulse reflection signal into a pre-trained convolutional neural network model, wherein the convolutional neural network model includes a signal waveform feature extractor 1, a gated screening adaptive attention module 2, a nonlinear association interaction module 3 and a classifier 4;

[0162] The signal waveform feature extractor 1 is used to extract signal waveform features from the optical pulse transmission signal and the optical pulse reflection signal respectively, and obtain a transmission signal waveform semantic feature map and a reflection signal waveform semantic feature map;

[0163] The gated screening adaptive attention module 2 is used to process the semantic feature map of the emission signal waveform and the semantic feature map of the reflection signal waveform to obtain the semantic saliency feature map of the emission signal waveform and the semantic saliency feature map of the reflection signal waveform;

[0164] The nonlinear association interaction module 3 is used to perform nonlinear association interaction processing on the semantically significant feature map of the transmitted signal waveform and the semantically significant feature map of the reflected signal waveform to obtain a semantically significant feature map of the transmitted signal-reflected signal waveform;

[0165] The classifier 4 is used to output the optical fiber winding recognition result according to the semantic association feature graph of the transmitted signal-reflected signal waveform.

[0166] Furthermore, the gated screening adaptive attention module 2 processes the semantic feature map of the emission signal waveform and the semantic feature map of the reflection signal waveform to obtain the semantic saliency feature map of the emission signal waveform and the semantic saliency feature map of the reflection signal waveform, including:

[0167] Performing multi-scale global pooling processing and position weighted sum on the semantic feature map of the transmitted signal waveform and the semantic feature map of the reflected signal waveform, respectively, to obtain a multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and a multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform;

[0168] The multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform are respectively subjected to bias processing to obtain a bias-adjusted multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and a bias-adjusted multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform;

[0169] The bias-adjusted transmitted signal waveform semantic association feature multi-scale pooling representation vector and the bias-adjusted reflected signal waveform semantic association feature multi-scale pooling representation vector are respectively activated through an activation function to obtain a transmitted signal waveform channel semantic association feature vector and a reflected signal waveform channel semantic association feature vector;

[0170] The transmission signal waveform channel semantic association feature vector and the reflection signal waveform channel semantic association feature vector are respectively input into the gating unit to perform masking and normalization operations on the feature values ​​at each position to obtain a normalized transmission signal waveform semantic weight vector and a normalized reflection signal waveform semantic weight vector;

[0171] The eigenvalues ​​of each position in the normalized transmitted signal waveform semantic weight vector and the normalized reflected signal waveform semantic weight vector are used as weighting coefficients, and weighted optimization is performed on each feature matrix along the channel dimension in the transmitted signal waveform semantic feature map and the reflected signal waveform semantic feature map to obtain the transmitted signal waveform semantic saliency feature map and the reflected signal waveform semantic saliency feature map.

[0172] Further, the semantic feature map of the transmitted signal waveform and the semantic feature map of the reflected signal waveform are respectively subjected to multi-scale global pooling processing and position weighted sum, to obtain a multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and a multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform, including:

[0173] Performing maximum-based global pooling processing, random-value-based global pooling processing, and average-based global pooling processing on the semantic feature graph of the transmitted signal waveform and the semantic feature graph of the reflected signal waveform, respectively, to obtain a transmitted signal waveform semantic global maximum pooling feature vector, a transmitted signal waveform semantic global random value pooling feature vector, a transmitted signal waveform semantic global average pooling feature vector, a reflected signal waveform semantic global maximum pooling feature vector, a reflected signal waveform semantic global random value pooling feature vector, and a reflected signal waveform semantic global average pooling feature vector;

[0174] The transmitted signal waveform semantic global maximum pooling feature vector, the transmitted signal waveform semantic global random value pooling feature vector and the transmitted signal waveform semantic global mean pooling feature vector are weighted according to position to obtain a multi-scale pooling representation vector of the transmitted signal waveform semantic association feature;

[0175] The reflected signal waveform semantic global maximum pooling feature vector, the reflected signal waveform semantic global random value pooling feature vector and the reflected signal waveform semantic global mean pooling feature vector are weighted according to position to obtain a reflected signal waveform semantic association feature multi-scale pooling representation vector.

[0176] Furthermore, the multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform are respectively subjected to bias processing, including:

[0177] The multi-scale pooling representation vector of the semantic association features of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association features of the reflected signal waveform are multiplied by the weight matrix respectively, and then added with the bias vector to obtain the bias-adjusted multi-scale pooling representation vector of the semantic association features of the transmitted signal waveform and the bias-adjusted multi-scale pooling representation vector of the semantic association features of the reflected signal waveform respectively.

[0178] Furthermore, the semantically associated feature vector of the transmitted signal waveform channel and the semantically associated feature vector of the reflected signal waveform channel are respectively input into the gating unit to perform masking and normalization operations on the feature values ​​at each position, including:

[0179] Based on the gating mechanism, the feature value of each position in the semantic association feature vector of the transmission signal waveform channel is compared with a preset threshold, the feature value greater than or equal to the preset threshold is taken as the original value, and the feature value less than the preset threshold is taken as 0, so as to obtain a masked transmission signal waveform channel semantic association feature vector;

[0180] Compare the feature value of each position in the semantic association feature vector of the reflection signal waveform channel with a preset threshold, take the original value for the feature value greater than or equal to the preset threshold, and take the feature value less than the preset threshold as 0, to obtain a masked reflection signal waveform channel semantic association feature vector;

[0181] Each eigenvalue in the masked transmission signal waveform channel semantic association feature vector and the masked reflection signal waveform channel semantic association feature vector is normalized to obtain a normalized transmission signal waveform semantic weight vector and a normalized reflection signal waveform semantic weight vector, respectively.

[0182] Furthermore, the nonlinear association interaction module 3 performs nonlinear association interaction processing on the semantically significant feature map of the transmitted signal waveform and the semantically significant feature map of the reflected signal waveform to obtain a semantically significant feature map of the transmitted signal-reflected signal waveform, including:

[0183] Interactively process the semantically saliency feature map of the transmitted signal waveform and the semantically saliency feature map of the reflected signal waveform to obtain a semantically interactive feature map of the transmitted signal and the reflected signal;

[0184] According to the transmission signal-reflection signal semantic interaction feature map and the preset adjustment hyperparameters, a transmission-reflection waveform semantic interaction nonlinear response value is calculated;

[0185] The semantic interaction nonlinear response value of the emission-reflection waveform is reshaped into a feature map to obtain the semantic association feature map of the emission signal-reflection signal waveform.

[0186] Furthermore, the semantically saliency feature map of the transmitted signal waveform and the semantically saliency feature map of the reflected signal waveform are divided position by position to obtain the transmitted signal-reflected signal semantic interaction feature map.

[0187] Further, according to the transmission signal-reflection signal semantic interaction feature map and the preset adjustment hyperparameters, the transmission-reflection waveform semantic interaction nonlinear response value is calculated, including:

[0188] Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by the first preset adjustment hyperparameter to obtain a first transmitted-reflected signal semantic interaction nonlinear factor quadratic term adjustment parameter;

[0189] Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by the second preset adjustment hyperparameter to obtain a first-order adjustment parameter of the nonlinear factor of the first transmitted-reflected signal semantic interaction;

[0190] Adding the first transmission-reflection signal semantic interaction nonlinear factor quadratic term adjustment parameter and the first transmission-reflection signal semantic interaction nonlinear factor linear term adjustment parameter to obtain the first transmission-reflection signal semantic interaction nonlinear factor;

[0191] Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by a third preset adjustment hyperparameter to obtain a second transmission-reflected signal semantic interaction nonlinear factor quadratic term adjustment parameter;

[0192] Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by a fourth preset adjustment hyperparameter to obtain a first-order adjustment parameter of a second transmitted-reflected signal semantic interaction nonlinear factor;

[0193] Adding the second transmission-reflection signal semantic interaction nonlinear factor quadratic term adjustment parameter to the second transmission-reflection signal semantic interaction nonlinear factor linear term adjustment parameter to obtain a second transmission-reflection signal semantic interaction nonlinear factor;

[0194] The first transmission-reflection signal semantic interaction nonlinear factor is divided by the second transmission-reflection signal semantic interaction nonlinear factor to obtain a transmission-reflection waveform semantic interaction nonlinear response value.

[0195] Furthermore, in some embodiments, the device further includes an early warning module for generating and displaying early warning information when the optical fiber winding identification result shows that the optical fiber is entangled.

[0196] The optical fiber winding detection method and device provided in the above embodiment have at least the following beneficial effects:

[0197] (1) Sending optical pulses through an optical time domain reflectometer, and collecting optical pulse emission signals and optical pulse reflection signals, and then introducing artificial intelligence-based signal processing and analysis algorithms to analyze and nonlinearly correlate the optical pulse emission signals and optical pulse reflection signals, so as to identify the implicit pattern and correlation feature information between the optical pulse emission signals and the optical pulse reflection signals, and thus determine whether the fiber entanglement problem occurs. In this way, the entanglement problem in the optical fiber link can be quickly identified and located, and the automatic identification of the fiber entanglement problem can be realized. When the fiber entanglement is detected, an entanglement warning prompt is generated, so that corresponding measures can be taken to effectively prevent the fiber entanglement;

[0198] (2) Through the processing of the gated filtering adaptive attention module, the features that are most critical for identifying fiber entanglement problems can be enhanced, while those less important information can be suppressed. In this way, the model can more clearly identify abnormal patterns in the signal. In other words, the gated filtering adaptive attention module can automatically adjust its focus according to the specific content of the signal. This dynamic selection capability enables the model to adapt to the various entanglement situations that may exist in the optical fiber link, which enables the model to better handle the personalized characteristics of different optical fiber links, so that the model can more accurately identify signs of optical fiber entanglement, thereby improving the accuracy of entanglement detection;

[0199] (3) Through the nonlinear correlation interaction module, the feature maps of the transmitted signal and the reflected signal can be fused. This fusion helps to capture the complex nonlinear relationship between the semantics of the two signals, thereby more comprehensively understanding the state of the optical fiber link and more accurately identifying the entanglement problem in the optical fiber link. In addition, the nonlinear correlation interaction of the semantic features of the transmitted signal waveform and the semantic features of the reflected signal waveform helps to improve the adaptability of the model to different types of optical fiber links, making the detection results more robust.

[0200] (4) During the model training process, by training the classifier through the relevant associated loss function value, it is possible to model the structural detail dependency relative to the macroscopic complexity in the classification scenario, thereby promoting the understanding of the macroscopic complex feature expression of the sample based on the structural details and improving the accuracy of the classification results obtained by the classifier-based fiber winding detector.

[0201] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for detecting optical fiber winding, characterized in that: include: Acquire the optical pulse transmission signal sent by the optical time domain reflectometer to the optical fiber to be detected, and receive the reflected optical pulse reflection signal; Inputting the optical pulse emission signal and the optical pulse reflection signal into a pre-trained convolutional neural network model, wherein the convolutional neural network model includes a signal waveform feature extractor, a gated screening adaptive attention module, a nonlinear association interaction module, and a classifier; The signal waveform feature extractor extracts signal waveform features from the optical pulse transmission signal and the optical pulse reflection signal respectively to obtain a transmission signal waveform semantic feature graph and a reflection signal waveform semantic feature graph; The gated screening adaptive attention module processes the semantic feature map of the emission signal waveform and the semantic feature map of the reflection signal waveform to obtain a semantic saliency feature map of the emission signal waveform and a semantic saliency feature map of the reflection signal waveform; The nonlinear association interaction module performs nonlinear association interaction processing on the semantically significant feature map of the transmitted signal waveform and the semantically significant feature map of the reflected signal waveform to obtain a semantic association feature map of the transmitted signal-reflected signal waveform; The classifier outputs an optical fiber winding recognition result according to the transmitted signal-reflected signal waveform semantic association feature map.

2. The method according to claim 1, characterized in that The gated screening adaptive attention module processes the semantic feature map of the emission signal waveform and the semantic feature map of the reflection signal waveform to obtain the semantic saliency feature map of the emission signal waveform and the semantic saliency feature map of the reflection signal waveform, including: The semantic feature map of the transmitted signal waveform and the semantic feature map of the reflected signal waveform are respectively subjected to multi-scale global pooling processing, position weighted sum, bias processing, activation processing, mask processing, normalization operation and weighted optimization to obtain the semantic saliency feature map of the transmitted signal waveform and the semantic saliency feature map of the reflected signal waveform.

3. The method according to claim 2, characterized in that The transmitted signal waveform semantic feature map and the reflected signal waveform semantic feature map are respectively subjected to multi-scale global pooling processing and position weighted sum, including: Performing global pooling processing based on maximum value, random value and average value on the semantic feature map of the transmitted signal waveform and the semantic feature map of the reflected signal waveform respectively, and obtaining a transmitted signal waveform semantic global maximum pooling feature vector, a transmitted signal waveform semantic global random value pooling feature vector, a transmitted signal waveform semantic global mean pooling feature vector, a reflected signal waveform semantic global maximum pooling feature vector, a reflected signal waveform semantic global random value pooling feature vector and a reflected signal waveform semantic global mean pooling feature vector respectively; The transmitted signal waveform semantic global maximum pooling feature vector, the transmitted signal waveform semantic global random value pooling feature vector and the transmitted signal waveform semantic global mean pooling feature vector are weighted according to position to obtain a multi-scale pooling representation vector of the transmitted signal waveform semantic association feature; The reflected signal waveform semantic global maximum pooling feature vector, the reflected signal waveform semantic global random value pooling feature vector and the reflected signal waveform semantic global mean pooling feature vector are weighted according to position to obtain a reflected signal waveform semantic association feature multi-scale pooling representation vector.

4. The method according to claim 3, characterized in that: After obtaining the multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform, bias processing is performed, including: The multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform are respectively multiplied by the weight matrix, and then added with the bias vector to obtain the bias-adjusted multi-scale pooling representation vector of the semantic association feature of the transmitted signal waveform and the bias-adjusted multi-scale pooling representation vector of the semantic association feature of the reflected signal waveform; After obtaining the bias-adjusted transmitted signal waveform semantic association feature multi-scale pooling representation vector and the bias-adjusted reflected signal waveform semantic association feature multi-scale pooling representation vector, activation processing is performed through an activation function to obtain the transmitted signal waveform channel semantic association feature vector and the reflected signal waveform channel semantic association feature vector.

5. The method according to claim 4, characterized in that After obtaining the semantically associated feature vector of the transmitted signal waveform channel and the semantically associated feature vector of the reflected signal waveform channel, they are input into the gating unit to perform masking and normalization operations on the feature values ​​at each position, including: Based on the gating mechanism, the feature value of each position in the semantic association feature vector of the transmission signal waveform channel is compared with a preset threshold, the feature value greater than or equal to the preset threshold is taken as the original value, and the feature value less than the preset threshold is taken as 0, so as to obtain a masked transmission signal waveform channel semantic association feature vector; Compare the feature value of each position in the semantic association feature vector of the reflection signal waveform channel with a preset threshold, take the original value for the feature value greater than or equal to the preset threshold, and take the feature value less than the preset threshold as 0, to obtain a masked reflection signal waveform channel semantic association feature vector; Normalizing each eigenvalue in the masked transmission signal waveform channel semantic association feature vector and the masked reflection signal waveform channel semantic association feature vector to obtain a normalized transmission signal waveform semantic weight vector and a normalized reflection signal waveform semantic weight vector, respectively; After obtaining the normalized transmitted signal waveform semantic weight vector and the normalized reflected signal waveform semantic weight vector, weighted optimization is performed to obtain a transmitted signal waveform semantic saliency feature map and a reflected signal waveform semantic saliency feature map.

6. The method according to claim 1, characterized in that The nonlinear association interaction module performs nonlinear association interaction processing on the semantically significant feature map of the transmitted signal waveform and the semantically significant feature map of the reflected signal waveform to obtain a semantically significant feature map of the transmitted signal-reflected signal waveform, including: Interactively process the semantically saliency feature map of the transmitted signal waveform and the semantically saliency feature map of the reflected signal waveform to obtain a semantically interactive feature map of the transmitted signal and the reflected signal; According to the transmission signal-reflection signal semantic interaction feature map and the preset adjustment hyperparameters, a transmission-reflection waveform semantic interaction nonlinear response value is calculated; The semantic interaction nonlinear response value of the emission-reflection waveform is reshaped into a feature map to obtain the semantic association feature map of the emission signal-reflection signal waveform.

7. The method according to claim 6, characterized in that The semantically saliency feature map of the transmitted signal waveform and the semantically saliency feature map of the reflected signal waveform are divided position by position to obtain the transmitted signal-reflected signal semantic interaction feature map.

8. The method according to claim 6, characterized in that According to the transmission signal-reflection signal semantic interaction feature map and the preset adjustment hyperparameters, the transmission-reflection waveform semantic interaction nonlinear response value is calculated, including: Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by the first preset adjustment hyperparameter to obtain a first transmitted-reflected signal semantic interaction nonlinear factor quadratic term adjustment parameter; Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by the second preset adjustment hyperparameter to obtain a first-order adjustment parameter of the nonlinear factor of the first transmitted-reflected signal semantic interaction; Adding the first transmission-reflection signal semantic interaction nonlinear factor quadratic term adjustment parameter and the first transmission-reflection signal semantic interaction nonlinear factor linear term adjustment parameter to obtain the first transmission-reflection signal semantic interaction nonlinear factor; Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by a third preset adjustment hyperparameter to obtain a second transmission-reflected signal semantic interaction nonlinear factor quadratic term adjustment parameter; Multiplying the square value of the feature value at each position in the transmitted signal-reflected signal semantic interaction feature map by a fourth preset adjustment hyperparameter to obtain a first-order adjustment parameter of a second transmitted-reflected signal semantic interaction nonlinear factor; Adding the second transmission-reflection signal semantic interaction nonlinear factor quadratic term adjustment parameter to the second transmission-reflection signal semantic interaction nonlinear factor linear term adjustment parameter to obtain a second transmission-reflection signal semantic interaction nonlinear factor; The first transmission-reflection signal semantic interaction nonlinear factor is divided by the second transmission-reflection signal semantic interaction nonlinear factor to obtain a transmission-reflection waveform semantic interaction nonlinear response value.

9. The method according to any one of claims 1 to 8, characterized in that: The method further comprises: When the optical fiber winding identification result shows that the optical fiber is entangled, a warning message is generated and displayed.

10. An optical fiber winding detection device applied to the method according to any one of claims 1 to 9, characterized in that: include: A signal acquisition module is used to acquire the optical pulse transmission signal sent by the optical time domain reflectometer to the optical fiber to be detected, and receive the reflected optical pulse reflection signal; A detection module, used for inputting the light pulse emission signal and the light pulse reflection signal into a pre-trained convolutional neural network model, wherein the convolutional neural network model includes a signal waveform feature extractor, a gated screening adaptive attention module, a nonlinear association interaction module and a classifier; The signal waveform feature extractor is used to extract signal waveform features from the optical pulse transmission signal and the optical pulse reflection signal respectively, to obtain a transmission signal waveform semantic feature graph and a reflection signal waveform semantic feature graph; The gated screening adaptive attention module is used to process the semantic feature map of the emission signal waveform and the semantic feature map of the reflection signal waveform to obtain the semantic saliency feature map of the emission signal waveform and the semantic saliency feature map of the reflection signal waveform; The nonlinear association interaction module is used to perform nonlinear association interaction processing on the semantically significant feature map of the transmitted signal waveform and the semantically significant feature map of the reflected signal waveform to obtain a semantic association feature map of the transmitted signal-reflected signal waveform; The classifier is used to output the optical fiber winding recognition result according to the transmitted signal-reflected signal waveform semantic association feature map.

Images