A phi-otdr vibration signal intelligent identification method for coastline security

Abstract

The application discloses a kind of coast security-oriented Φ-OTDR vibration signal intelligent identification method, based on mixed expert network and cross attention mechanism realizes the positioning and identification of invasion signal.For the strong tidal noise interference, signal attenuation difference is big and the problem that hidden invasion target identification is difficult in coastline perimeter monitoring, the application constructs deep learning model based on MoE architecture, utilizes the energy distribution and statistical characteristics of input signal perceived by gate network, dynamically activates specific convolution module in expert library for weak invasion, strong interference or background noise, to realize the adaptive extraction of vibration mode under different sea conditions;At the same time, introduce the pair cross attention mechanism to deep interaction of multi-scale features, enhance the capture ability of weak invasion signal under complex background noise;Combined with task alignment learning to realize high-precision positioning and classification, significantly improve the anti-interference ability and invasion event identification accuracy of Φ-OTDR system in complex coastal environment.

Description

Technical Field

[0001] This invention relates to the fields of fiber optic sensing and artificial intelligence signal processing technology, and in particular to a method for intelligent identification of vibration signals using a Φ-OTDR for coastal security. Background Technology

[0002] Phase-sensitive optical time-domain reflectometers (Φ-OTDRs) have become an important technical means for perimeter security along coastlines due to their advantages such as long-distance and passive monitoring. However, the coastal environment is extremely complex. Waves, tides, and sea breezes generate continuous non-stationary background noise, which can easily drown out weak intrusion signals. Furthermore, the coastline includes various geological formations such as mudflats, sandy areas, and reefs, and optical cables are buried in different media with drastically different vibration transmission characteristics. In addition, the targets detected by the model are diverse, ranging from strong signals of mechanical excavation and ships approaching the shore to extremely subtle signals of divers coming ashore and crawling forward.

[0003] To address these challenges, existing research has proposed various deep learning solutions, but due to the static nature of the network structure, they are still insufficient to meet the requirements for accurate detection in complex sea conditions.

[0004] First, Reference 1 (Shi Yi*; Wang Yuanye; Zhao Lei; Fan, Zhun . An EventRecognition Method for Φ-OTDR Sensing System Based on Deep Learning[J]. Sensors, 2019, 19, 3421) treats the spatiotemporal matrix as an image and proposes a recognition method based on a two-dimensional convolutional neural network (2D-CNN). However, the limitations of this method stem from its multiple max-pooling operations and global weight sharing mechanism. While reducing dimensionality, the pooling layer irreversibly loses the high-frequency details of weak signals. Furthermore, the fixed convolutional kernel weights cannot dynamically adjust the filtering parameters according to changes in the intensity of background noise, causing the network to easily misjudge non-stationary broadband noise as intrusion signals when waves are crashing violently.

[0005] Secondly, Reference 2 (Huijuan Wu, Jiping Chen, Xiangrong Liu, Yao Xiao, Mengjiao Wang, Yi Zheng, and Yunjiang Rao, "One-Dimensional CNN-Based Intelligent Recognition of Vibrations in Pipeline Monitoring With DAS," J. Lightwave Technol. 37, 4359-4366 (2019)) proposed an intelligent recognition network based on multi-scale one-dimensional convolution (Multi-scale 1D-CNN). Although its parallel multi-size convolutional kernels can cover signal features of different frequencies, this structure fails to solve the masking effect of signal energy dimension. Because the network tends to fit high-amplitude samples that contribute greatly to the loss function during training, such as mechanical mining, the convolutional kernel weights are dominated by strong signals. During the inference stage, when faced with weak frogman signals, even if the small-scale convolutional kernels can capture their frequency features, their extremely low feature response values ​​will be overwhelmed by the high response values ​​of background noise or strong interference signals in subsequent pooling layers, resulting in missed detections.

[0006] Recently, reference 3 (Yi Shi, Jiewei Chen, Shangwei Dai, Xinyu Liu, and ChuliangWei, "Multi-signal feature fusion method with an attention mechanism for the Φ-OTDR event recognition system," Opt. Express 30, 42086-42096, 2022) introduced an attention mechanism to optimize multi-feature fusion. However, this method is essentially based on a self-attention mechanism. That is, the query (Q), key (K), and value (V) in its attention calculation all originate from the output of the same feature scale. This feature homology leads to the model lacking a global perspective across scales. When a feature layer is disturbed by strong ocean noise, the noise components present in both Q and K will produce false high correlations, resulting in the noise obtaining a high response value. Due to the lack of deep semantic features as key-value pairs K / V to perform cross-scale cross-validation of shallow detail features Q, this model cannot effectively suppress non-stationary background noise while preserving weak signal details.

[0007] Therefore, there is an urgent need for an intelligent processing method that can dynamically adjust the network structure according to sea conditions and signal characteristics, and can adaptively focus on key information. Summary of the Invention

[0008] The problem to be solved by this invention is to provide an intelligent identification method for Φ-OTDR vibration signals for coastal security. Based on a hybrid expert network, the method introduces the MoE mechanism to realize dynamic routing of feature extraction and uses the cross-attention mechanism to enhance feature fusion, thereby improving the detection performance of the system in complex environments.

[0009] This invention adopts the following technical solution: a Φ-OTDR vibration signal intelligent identification method for coastal security, comprising the following steps:

[0010] Step 1: Data Preparation and Preprocessing: External vibration signals are collected using the Φ-OTDR system, and the signals are normalized and manually labeled to determine the specific categories of vibration events and their temporal and spatial locations, resulting in a spatiotemporal matrix sample dataset containing category and location labels. The empty matrix sample dataset contains training and validation sets, providing a foundation for subsequent model training.

[0011] Step 2: Construct a fiber optic vibration signal localization and classification model based on the Mixture of Experts (MoE) architecture, including: the MoE backbone network, the Pairwise Cross-Attention PAN network, and the detection network.

[0012] Step 3: Input the spatiotemporal matrix into the MoE backbone network for feature extraction; the MoE backbone network is divided into a basic transformation stage, a main feature extraction stage, and a MoE enhancement stage;

[0013] The data first enters the basic transformation stage (Stage 1) to complete the basic spatiotemporal feature mapping; then, the data flow is divided into the main feature extraction path and the MoE enhancement path, which enter the main feature extraction stage (Stage 2) and the MoE enhancement stage (Stage 3) respectively.

[0014] In the MoE enhancement path, the statistical features of the input signal are calculated by a gating network, the top-K experts are dynamically selected from the expert database to process the features, and the expert outputs are used as attention weights to modulate the main features output by the main feature extraction path, resulting in three enhanced feature maps at different scales.

[0015] Step 4: Input the three enhanced feature maps of different scales output by the MoE backbone network into the paired cross attention path aggregation network. By pairing feature maps of different scales, feature interaction and fusion are performed using the multi-head cross attention mechanism to output the enhanced feature pyramid.

[0016] Step 5: Input the enhanced feature pyramid into the detection network to predict and adjust the localization parameters and classification probabilities, and combine the task alignment learning (TAL) loss function to train and optimize the model.

[0017] As a preferred embodiment, the expert database in the MoE backbone network includes two types of experts:

[0018] Basic experts are convolutional processing units independently set for each spatial node of the input feature, with the number of units matching the number of spatial nodes, used to extract local fine features;

[0019] Specialized experts include large convolution kernel experts for capturing weak signal features, ultra-large convolution kernel experts for capturing strong signal features, and normalization experts for processing high-frequency noise, as well as experts for processing global and special pattern signals.

[0020] Specifically, there are 100 basic experts and 3 specialized experts.

[0021] As a preferred embodiment, the gated network of the MoE backbone network processes the following:

[0022] It receives input features, including node-level spatiotemporal features and global statistical features (including average energy, global variance, maximum energy, and minimum energy), and forms a routing vector;

[0023] Based on the routing vector, the activation weight of each expert in the expert database is calculated through the fully connected layer, and the K experts with the highest weights are dynamically selected to participate in the calculation.

[0024] After the activated experts process the signal, the output is weighted and aggregated and activated by Sigmoid to generate an attention weight map.

[0025] The attention weight map is multiplied element-wise with the main features output by the main feature extraction path to output an enhanced feature map with adaptive properties.

[0026] As a preferred embodiment, the pairwise cross-attention path aggregation network processes the following:

[0027] The three input enhanced feature maps at different scales are flattened in space and mapped through channels, and then uniformly mapped to the same embedding dimension.

[0028] The three feature maps are combined in pairs, and bidirectional multi-head cross-attention is calculated for each pair to generate fused features;

[0029] By utilizing learnable residual weights, the original projected features obtained after basic spatiotemporal feature mapping are weighted and aggregated with the corresponding cross-attention fusion features, and then restored back to the original number of channels and spatial dimensions to form a high-level feature pyramid containing rich contextual information.

[0030] As a preferred embodiment, the detection network includes a detection head that predicts localization adjustment parameters and a detection head that predicts classification probabilities, used to generate adjustment parameters and prediction probabilities for default anchor boxes for each prediction point.

[0031] The number of each detection head is the same as the number of layers in the multi-level enhanced feature map; each detection head contains two sets of convolutional layers, which are used to predict the bounding box regression parameters and class confidence of each anchor box, respectively.

[0032] As a preferred embodiment, the loss function adopts task alignment focal length loss, introduces alignment metric, calculates the intersection-union ratio of the predicted box and the ground truth box as a measure of localization quality (IoU), and uses the product of classification score and localization quality as the soft label of positive sample, dynamically adjusting the gradient contribution of classification and localization tasks.

[0033] As a preferred embodiment, the spatiotemporal matrix sample has an input dimension of [Batch, 1, T, S]; where Batch represents the batch size of samples, 1 represents a single channel, T represents the time series length, and S represents the number of spatial sensing nodes.

[0034] The MoE backbone network compresses the temporal dimension T to one dimension while keeping the spatial dimension S unchanged, and outputs three enhanced feature maps of different scales with dimensions [Batch, 1, T, W]. The width W takes values ​​of 20, 40, and 60 respectively, with different widths corresponding to differentiated receptive fields.

[0035] The MoE backbone network compresses the time dimension T to one dimension while keeping the spatial dimension S unchanged.

[0036] As a preferred embodiment, the fiber optic vibration signal localization and classification model further includes data post-processing, specifically including:

[0037] The prediction adjustment parameters are applied to the default length to obtain the prediction results; low-probability prediction results are filtered out; prediction results exceeding the input sample size are pruned; and prediction results of the same category are processed by NMS.

[0038] During the training process of the model, loss functions are set separately for localization and classification tasks. The total loss function is a linear superposition of the loss functions for localization and classification tasks. Iterative training is performed using stochastic gradient descent.

[0039] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0040] 1. The vibration signal localization and identification method of the Φ-OTDR system of the present invention is based on a hybrid expert network and a cross-attention mechanism, which can simultaneously achieve high-precision localization and identification of complex external vibration signals, and can significantly improve the adaptability, intelligence and processing efficiency of the Φ-OTDR system in complex environments.

[0041] 2. The pairwise cross-attention path aggregation network constructed in this invention performs pairwise deep interaction and fusion of feature maps at different scales through a multi-head attention mechanism. This can uncover the complementary relationship between deep semantic information and shallow detail features, enabling the model to more sensitively capture weak signal features that are difficult to distinguish, thereby enhancing the accuracy of localization and the reliability of classification.

[0042] 3. The hybrid expert backbone network proposed in this invention introduces a dynamic gating routing mechanism to replace the traditional static feature extraction. It can adaptively activate the most targeted expert modules according to the energy and statistical distribution characteristics of the input signal, which can effectively solve the problem of imbalance between strong and weak signal features in the Φ-OTDR raw data. Attached Figure Description

[0043] Figure 1 This is a spatiotemporal waterfall diagram of vibration events according to an embodiment of the present invention;

[0044] Figure 2 This is a diagram of the optical fiber vibration signal localization and classification model architecture of the present invention;

[0045] Figure 3 This is a training loss curve diagram of an embodiment of the present invention;

[0046] Figure 4 This is an averaged mAP curve from an embodiment of the present invention;

[0047] Figure 5 These are F1 score graphs under different prediction probability thresholds in embodiments of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0049] In one embodiment of the present invention, taking a coastal border defense perimeter monitoring and early warning system as an example, the vibration signal localization and identification method of the Φ-OTDR system of the present invention is applied to locate and identify vibration signals under complex sea conditions.

[0050] S1: Data Acquisition and Scenario Definition.

[0051] The Φ-OTDR sensing optical cable was laid along the coastline, traversing beaches, mudflats, and shallow sea areas. Background data under different tidal conditions, as well as simulated intrusion data, were collected. Based on actual security needs, vibration events were categorized as: "frogman landing," "ship landing," "knocking," and "digging." Figure 1 The spatiotemporal waterfall diagrams show that different categories of events have distinct morphological differences.

[0052] S2. Divide the labeled data into equal-sized segments to construct a spatiotemporal matrix sample dataset, which includes a validation set and a test set.

[0053] In this embodiment, the spatiotemporal matrix sample has an input dimension of [Batch, 1, T, S], where Batch represents the batch size of samples, 1 represents a single channel, T represents the time series length, and S represents the number of spatial sensing nodes.

[0054] S3. Construct a fiber optic vibration signal localization and classification model based on sample size. The overall network architecture is as follows: Figure 2 As shown, the model mainly consists of four parts: the MoE backbone network, the pairwise cross-attention feature fusion network, the detection network, and post-processing.

[0055] 1. In the MoE backbone network, this embodiment introduces a hybrid expert mechanism, and the network is divided into three phases:

[0056] First, in the first stage, the input signal undergoes preliminary spatiotemporal feature transformation through grouped convolution, converting the original high-dimensional input into a basic feature map, thus completing the basic spatiotemporal feature mapping.

[0057] Subsequently, the network is divided into two paths: one is the main feature extraction path based on ordinary convolution (main feature extraction path), which is used to extract general features of the signal; the other is the attention enhancement path based on MoE (MoE enhancement path).

[0058] The core logic of the MoE-enhanced pathway is described as follows:

[0059] ;

[0060] ;

[0061] ;

[0062] in, It is a routing vector composed of statistical measures such as the mean and variance of the node-level energy of the input features; , These represent the local feature vector and the global statistical feature vector of the input signal in the spatial node dimension, respectively. Indicates the concatenation function;

[0063] It is a gated network used to calculate the activation weights of each expert in the expert database. For normalization function, This represents the normalized activation probability distribution vector of each expert module under the current input;

[0064] This means that only the K experts with the highest weights are selected for the calculation. This represents the selected expert module. Indicates the selected number Normalized activation weights corresponding to each expert module This represents the enhanced spatiotemporal feature map output after weighted feature fusion by multiple expert modules.

[0065] Specifically, in this embodiment, the expert database includes 100 basic experts for each spatial node, and 3 specialized experts for weak signals, strong signals, and noise. The output of the MoE path is processed by an activation function to generate an attention weight map, which is then multiplied element-wise with the main feature map of Stage 2. This allows for adaptive enhancement of the backbone features using the specific pattern knowledge learned by the MoE, resulting in a multi-level enhanced feature map.

[0066] In this embodiment, the multi-level enhanced feature map output by the MoE backbone network has three dimensions [Batch, 1, TW], where the width W is 20, 40, and 60 respectively, with different widths corresponding to different receptive fields.

[0067] 2. Feature fusion is performed in a pairwise cross-attention feature fusion network.

[0068] Furthermore, to improve the accuracy of vibration signal classification and localization, a pairwise cross-attention network (PAN) is used for feature fusion between feature pyramids. For the predicted feature map, feature maps at different levels contain complementary semantic and detailed information. Traditional simple stacking cannot fully utilize this complementarity.

[0069] In this process, firstly, the feature maps of the three different scales output by the backbone network are uniformly projected to the same embedding dimension through linear mapping and flattened in space.

[0070] Next, the three feature maps are paired up. For each pair of feature combinations, a multi-head attention mechanism is used to calculate bidirectional cross-attention, thereby capturing deep correlations across scales. For example, in the combination of feature 0 and feature 1, attention is calculated with feature 0 as the query and feature 1 as the key and value, respectively, and attention is calculated with feature 1 as the query and feature 0 as the key and value, thereby capturing complementary information between features at different scales.

[0071] Finally, using learnable residual weight parameters, the original projected features are weighted and aggregated with all their corresponding cross-attention fusion results, restoring the original spatial dimension. The aggregated features are then back-projected to restore the original number of channels and spatial structure, forming a high-level feature pyramid containing rich contextual information. This design enables the model to automatically focus on the feature regions most valuable for localization and classification.

[0072] 3. In the detection network, the network generates adjustment parameters and prediction probabilities for the default anchor box for each prediction point.

[0073] Both tasks are accomplished using two one-dimensional convolutional layers with a kernel size of 3. The number of output channels used for classification and localization corresponds to the number of object categories and the tuning parameters, respectively. Each point in the predicted feature map is used for prediction on an anchor box with a default length of three widths. First, to simplify the location prediction task, a set of default widths is established for each prediction point in the predicted feature map. The default width depends on the receptive field of the feature map.

[0074] The detection network consists of detection heads that predict localization and adjust parameters and detection heads that predict classification probabilities. The number of each type of detection head is the same as the number of layers in the multi-level feature map.

[0075] 4. Perform post-processing.

[0076] In this embodiment, the post-processing mainly performs four tasks: 1) applying the prediction adjustment parameters to the default length to obtain the prediction result; 2) filtering low-probability prediction results; 3) pruning prediction results that exceed the input sample size; and 4) performing NMS processing on prediction results of the same category.

[0077] S4. Perform model training.

[0078] During model training, this embodiment employs Task Alignment Focal Loss (TAL Focal Loss) for both localization and classification tasks. The total loss function is not a simple linear summation but introduces an alignment metric, using the product of the classification score and the localization quality (IoU) as a soft label for positive samples. This dynamically adjusts the gradient contributions for both classification and localization tasks, and iterative training is performed using stochastic gradient descent.

[0079] S5. Train the model using the training dataset, and then validate it using the validation set after training.

[0080] Based on the actual situation of the dataset, the model was finally selected to output a maximum of 10 predicted targets in the prediction process, according to the prediction probability from high to low. The number of correct matches TP (TruePositive), the number of missed detections FN (False Negative), and the number of wrong (repeated) detections FP (False Positive) were calculated by comparing the predicted targets with the real sample labels.

[0081] Two factors affect matching: the prediction probability threshold and the IOU (Intersection over Union) threshold between the predicted length and the actual length. That is, assuming the predicted category is correct, a match is considered successful only if both the predicted probability threshold and the IOU between the predicted length and the actual length exceed a certain threshold.

[0082] Furthermore, precision and recall can be calculated, which are related to the nuisance alarm rate (NAR) and the missed alarm rate (MAR) in pipeline leak monitoring tasks, respectively.

[0083] ;

[0084] .

[0085] In practical applications, the model must consider both the false positive rate and the false negative rate. Therefore, the F1-score can be obtained by combining precision and recall.

[0086] ;

[0087] Similarly, by using percision and recall as the horizontal and vertical axes, we obtain the PR curve. The area under the curve is AP (Average Precision), and the average AP for all categories is mAP (Mean Average Precision).

[0088] Assuming the IOU threshold for successful matches is set to 0.5, and the prediction probability threshold consists of 100 values ​​ranging from 0 to 1 with intervals of 0.01, the trained network is validated and evaluated using a validation set after each training iteration throughout the training process. The AP under different conditions is then summed and averaged across the 100 probability thresholds. Finally, the training loss versus the averaged mAP curve is obtained, as shown below. Figure 3 and Figure 4 As shown.

[0089] The curve shows that the mAP reaches a high value of 97.5%, indicating that the proposed method has high precision and recall, while the false positive and false negative rates are very low. Furthermore, with the preset TP condition of IOU=0.5, the F1 score under different prediction probability thresholds is plotted as follows: Figure 5 As shown.

[0090] Based on the F1 score curve, the IOU for prediction was set to 0.5, and data with a confidence level below 0.5 were filtered out. The confusion matrix shown in Table 1 was plotted, and the percision and recall were calculated as shown in Table 2.

[0091] Table 1: Confusion Matrix

[0092]

[0093] Table 2: Test Results Table

[0094]

[0095] As can be seen, the model of this invention exhibits excellent feature extraction accuracy and class decoupling ability with both IOU and confidence level set at 0.5. From the perspective of action recognition, the model achieved zero false negatives (100% recall) for typical human actions such as "tapping" and "digging," fully demonstrating the high reliability of the algorithm in capturing dynamic behavioral features. Particularly noteworthy is that no cross-class confusion occurred in this test, indicating that the model achieved "zero confusion" classification, strongly verifying the strong orthogonality of the feature extraction network at the semantic level.

[0096] Despite the presence of a small number of false alarms in the "ship landing" category due to the complex wave texture of the sea surface (precision rate of approximately 98.04%), the model was still able to reliably capture the main target. For the more concealed "frogman landing" small target, the model still achieved 100% recall in the complex beach background, while keeping the number of false detections at an extremely low level, accurately detecting the vast majority of real samples, and fully meeting the detection needs in actual engineering scenarios.

[0097] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent identification of vibration signals using a Φ-OTDR for coastal security, characterized in that, Includes the following steps: Step 1: Use the Φ-OTDR system to collect external vibration signals, perform normalization preprocessing and labeling on the signals, determine the specific category of the vibration event and its temporal and spatial location, and obtain a spatiotemporal matrix sample dataset containing category and location labels. Step 2: Construct a fiber optic vibration signal localization and classification model based on a hybrid expert architecture, including: MoE backbone network, paired cross-attention path aggregation network, and detection network; Step 3: Input the spatiotemporal matrix into the MoE backbone network. After basic spatiotemporal feature mapping, the data stream enters the main feature extraction path and the MoE enhancement path respectively. The MoE enhancement path calculates the statistical features of the input signal through a gating network, dynamically selects Top-K experts from the expert database to process the features, and uses the expert output as attention weights to modulate the main features output by the main feature extraction path, outputting three enhanced feature maps at different scales. Step 4: Input the three enhanced feature maps of different scales output by the MoE backbone network into the pairwise cross-attention path aggregation network. By pairing feature maps of different scales together, feature interaction and fusion are performed using the multi-head cross-attention mechanism to output the enhanced feature pyramid. Step 5: Input the enhanced feature pyramid into the detection network to predict and adjust the localization parameters and classification probabilities, and train and optimize the model by combining the task alignment learning loss function.

2. The Φ-OTDR vibration signal intelligent recognition method according to claim 1, characterized in that, In the MoE backbone network, the expert database contains two types of experts: Basic Expert: Convolutional processing units are set up independently for each spatial node of the input feature, with the number consistent with the number of spatial nodes, and are used to extract local fine features; Specialized experts include large convolution kernel experts for capturing weak signal features, ultra-large convolution kernel experts for capturing strong signal features, and normalization experts for processing high-frequency noise, as well as experts for processing global and special pattern signals.

3. The Φ-OTDR vibration signal intelligent recognition method according to claim 1, characterized in that, The gated network of the MoE backbone network processes the following: It receives input features, including node-level spatiotemporal features and global statistical features, and forms a routing vector; Based on the routing vector, the activation weight of each expert in the expert database is calculated through the fully connected layer, and the K experts with the highest weights are dynamically selected to participate in the calculation. After the activated experts process the signal, the output is weighted and aggregated and activated by Sigmoid to generate an attention weight map. The attention weight map is multiplied element-wise with the main features output by the main feature extraction path to output an enhanced feature map with adaptive properties.

4. The Φ-OTDR vibration signal intelligent recognition method according to claim 1, characterized in that, The pairwise cross-attention path aggregation network processes the following: The three input enhanced feature maps at different scales are flattened in space and mapped through channels, and then uniformly mapped to the same embedding dimension. The three feature maps are combined in pairs, and bidirectional multi-head cross-attention is calculated for each pair to generate fused features; By utilizing learnable residual weights, the original projected features obtained after basic spatiotemporal feature mapping are weighted and aggregated with the corresponding cross-attention fusion features, and then restored back to the original number of channels and spatial dimensions to form a high-level feature pyramid containing rich contextual information.

5. The Φ-OTDR vibration signal intelligent recognition method according to claim 1, characterized in that, The detection network includes a detection head that predicts localization adjustment parameters and a detection head that predicts classification probabilities, used to generate adjustment parameters and prediction probabilities for default anchor boxes for each prediction point.

6. The Φ-OTDR vibration signal intelligent recognition method according to claim 6, characterized in that, The number of each detection head is the same as the number of layers in the multi-level enhanced feature map; each detection head contains two sets of convolutional layers, which are used to predict the bounding box regression parameters and class confidence of each anchor box, respectively.

7. The Φ-OTDR vibration signal intelligent recognition method according to claim 6, characterized in that, The loss function adopts task alignment focal length loss, introduces alignment metric, calculates the intersection-union ratio of predicted box and ground truth box as a measure of localization quality, and uses the product of classification score and localization quality as soft label for positive sample, dynamically adjusting the gradient contribution of classification and localization tasks.

8. The Φ-OTDR vibration signal intelligent recognition method according to claim 1, characterized in that, The spatiotemporal matrix sample has an input dimension of [Batch, 1, T, S]; where Batch represents the batch size of samples, 1 represents a single channel, T represents the time series length, and S represents the number of spatial sensing nodes. The MoE backbone network compresses the temporal dimension T to one dimension while keeping the spatial dimension S unchanged, and outputs three enhanced feature maps of different scales with dimensions [Batch, 1, T, W]. The width W takes values ​​of 20, 40, and 60 respectively, with different widths corresponding to differentiated receptive fields.

9. The Φ-OTDR vibration signal intelligent recognition method according to claim 1, characterized in that, The fiber optic vibration signal localization and classification model also includes data post-processing, specifically including: The prediction adjustment parameters are applied to the default length to obtain the prediction result; Filter out low-probability prediction results; Prune predictions that exceed the input sample size, and perform NMS processing on predictions of the same category.

Images