Vibration signal positioning and identification method oriented to phi-OTDR (Optical Time Domain Reflectometer) system

Abstract

The invention discloses a vibration signal positioning and identification method for a phi-OTDR (Optical Time Domain Reflectometer) system, which comprises the following steps of: firstly, acquiring a distributed optical fiber sensing signal, then segmenting the distributed optical fiber sensing signal, and marking a signal category and a signal generation position to obtain a space-time matrix sample data set; dividing the space-time matrix sample data set to obtain a training set and a verification set; constructing an optical fiber vibration signal positioning and classification model according to the size of a sample, inputting a training set to train the model, inputting a verification set to the trained model to verify the model, then carrying out model hyper-parameter adjustment to adapt to the field condition, and finally completing the classification and positioning of external vibration signals. According to the invention, the intelligence of the phi-OTDR system can be improved, the processing efficiency of the phi-OTDR is effectively improved, and the labor burden is reduced.

Description

Technical Field

[0001] The invention belongs to the technical field of optical fiber sensing, and in particular relates to a vibration signal positioning and identification method for a Φ-OTDR system. Background Art

[0002] Φ-OTDR technology has been used for acoustic and vibration measurement of the Internet of Things and for building smart cities. With its long measurement distance and continuous anti-interference detection method, it can measure vibration events at any position along the optical fiber and restore the amplitude and phase of external vibration with high fidelity. Therefore, it has been widely used in perimeter security and power system monitoring, oil and gas energy exploration, pipeline leakage detection and other fields, providing a high-sensitivity, large-capacity, low-cost long-distance all-weather acoustic / vibration dynamic detection technology solution for the urban Internet of Things.

[0003] In the application of distributed fiber optic acoustic field sensing, fully automatic signal positioning and recognition are two important tasks that affect the intelligence level of the system. Most signal processing methods are generally positioning methods or recognition methods, or simply adding the two methods. For example, "Mixed Intrusion Events Recognition Based on Group Convolutional Neural Networks in DAS System" uses a classification method to solve the classification problem of multiple vibration events in a sample, but it is unable to accurately locate each vibration event. "PIG Tracking Utilizing Fiber OpticDistributed Vibration Sensor and YOLO" accurately locates the location of the vibration event, but does not achieve the classification task. At the same time, due to the diversity of intrusion methods and the complexity of environmental noise, the positioning and recognition of vibration signals have high interference alarms, and the above methods are difficult to efficiently achieve positioning and recognition tasks. Summary of the invention

[0004] Technical problem to be solved: The present invention proposes a vibration signal positioning and identification method for a Φ-OTDR system, which can simultaneously realize the positioning and identification of external vibration signals, effectively improve the intelligence level and processing efficiency of the Φ-OTDR system, and reduce the manual burden.

[0005] Technical solution:

[0006] A vibration signal positioning and identification method for a Φ-OTDR system, the vibration signal positioning and identification method comprising the following steps:

[0007] Step 1: Use the Φ-OTDR system to collect external vibration signals, and mark the collected vibration signals with their categories and occurrence locations;

[0008] Step 2: Split the labeled vibration signal data into fixed-size data and construct a space-time matrix sample data set. The space-time matrix sample data set includes a validation set and a training set. All samples in the data set are space-time matrices of the same size, and each sample has a category label and a vibration position label.

[0009] Step 3, construct a fiber optic vibration signal positioning and classification model based on the sample size; the fiber optic vibration signal positioning and classification model includes four parts: a backbone network, a feature extraction and attention fusion network, a detection network and a post-processing structure; the backbone network uses multiple group convolutions to superimpose, and connects multiple residual modules in series after the group convolution to perform preliminary feature extraction and compression on the input vibration signal, and obtain multi-layer feature maps with different compression degrees, and output them to the feature extraction and attention fusion network; the feature extraction and attention fusion network uses the FPN structure to further extract features from the multi-layer feature maps, and combines the bottom-level features extracted by the backbone network with the FPN structure. The feature maps of different levels extracted by the N structure are fused, and a one-dimensional multi-layer compressed feature map is output as a predicted feature map and sent to the detection network; the detection network includes a first detection head for predicting positioning adjustment parameters and a second detection head for predicting classification probability, which respectively predict positioning adjustment parameters and classification probability, and the number of each detection head is the same as the number of layers of the multi-level feature map; the post-processing structure applies the positioning adjustment parameters to the preset length to obtain the signal position, filters the low-probability prediction results, sets a confidence threshold, and filters out the detection frames below the threshold. The threshold is usually set to 0.75 in field applications, and the detection frames below 0.75 will be discarded. The prediction results that exceed the input sample size are cropped, and the prediction results of the same category are processed by non-maximum suppression (NMS) to eliminate redundant detection frames;

[0010] Step 4: Use the training set to train the optical fiber vibration signal positioning and classification model. After the training is completed, use the validation set for verification and optimize the model hyperparameters based on the verification results.

[0011] Furthermore, in step 1, the categories of the vibration signal include at least five categories: touching the optical fiber, walking, manual digging, excavator digging and climbing.

[0012] Furthermore, in step 2, the backbone network uses multiple group convolution blocks G-BLOCK to perform preliminary feature extraction and compression on the input vibration signal, and the process of obtaining multi-layer feature maps with different compression degrees includes the following steps:

[0013] The input vibration signal is divided into n groups, and the convolution kernels of n group convolution blocks G-BLOCK extract features from each group. The convolution kernel of each group convolution block G-BLOCK performs corresponding convolution on only one of the input groups; the group convolution stage is described as follows:

[0014]

[0015] Where O is the output, f is the nonlinear activation function that introduces nonlinear factors, which is used to increase the nonlinear fitting ability of the model. C, w_i, b_i, x_Parti represent the number of convolution kernels, convolution weights, convolution bias, and the input of the i-th group respectively.

[0016] G-BLOCK is used for initial feature extraction to learn the vibration characteristics of different vibration positions and to extract features in the time dimension while maintaining the invariance of the spatial dimension;

[0017] An FP-Block consisting of four RES-Blocks stacked together is added to generate the original feature pyramid to improve the correlation between the outputs of the group convolution and obtain predicted feature maps with different receptive fields; among them, the signal with a larger vibration range is predicted by a small feature map with a large receptive field, and the signal with a narrower vibration range is predicted by a large feature map with a smaller receptive field.

[0018] Furthermore, in step 2, the process in which the feature extraction and attention fusion network outputs a one-dimensional multi-layer compressed feature map as a prediction feature map and sends it to the detection network includes the following steps:

[0019] The feature maps of each layer are processed by 2DConv, and then the feature maps of the upper layer are upsampled and added to the similar feature maps of the lower layer, so that the target position and high semantic information are fused together;

[0020] Add multiple g-conv layers, each of which compresses each fused feature map, compresses all output feature maps in the time dimension, and converts each two-dimensional feature map into a one-dimensional feature map;

[0021] In the Fusion-A module, all compressed one-dimensional feature maps are fused with the original 2D feature map of the lowest layer feature map to enhance the signal positioning effect; the Fusion-A module contains a g-conv layer for compressing the original feature map and two 1×1 convolution layers for adjusting the number of channels. A one-dimensional multi-layer compressed feature map is obtained by dot multiplication of the outputs of the two 1×1 convolution layers, which is output to the detection network as a predicted feature map.

[0022] Further, in step 2, the detection network includes a first detection head for predicting the positioning adjustment parameter and a second detection head for predicting the classification probability, respectively predicting the positioning adjustment parameter and the classification probability

[0023] A set of default length values ​​is established at each prediction point of the predicted feature map, and the default length depends on the receptive field of the feature map; each point of the predicted feature map is used to make predictions on anchor boxes with three default lengths;

[0024] Two 1D convs of size 3 are used to generate the adjustment parameters and prediction probabilities of the anchor boxes for each predicted point. The number of output channels of the 1D convs used for classification and positioning is the number of object categories and the number of adjustment parameters, respectively.

[0025] Furthermore, in step 4, during the training process of the optical fiber vibration signal positioning and classification model, loss functions are set for the positioning and classification tasks respectively, and the total loss function is the linear superposition of the positioning task loss function and the classification task loss function. The random gradient descent method is used to iteratively train the optical fiber vibration signal positioning and classification model.

[0026] Beneficial effects:

[0027] First, the vibration signal positioning and identification method for the Φ-OTDR system of the present invention proposes a feature fusion module based on the attention mechanism, which enables the model to focus on the positioning of signals that are difficult to distinguish, thereby enhancing the accuracy of positioning.

[0028] Second, the vibration signal positioning and identification method for the Φ-OTDR system of the present invention adopts group convolution to perform initial feature extraction, learns the characteristics of vibration changes at different vibration positions, and realizes feature extraction in the time dimension, thereby realizing data processing at high speed and efficiently, and then completes the feature extraction of the initial imbalance data.

[0029] Thirdly, the vibration signal positioning and identification method for the Φ-OTDR system of the present invention realizes the positioning and identification of external vibration signals at the same time, thereby improving the intelligence level and processing efficiency of the Φ-OTDR system. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a block diagram of the optical fiber vibration signal positioning and classification model of the present invention;

[0031] Figure 2 It is a parameter diagram of the optical fiber vibration signal positioning and classification model of the present invention;

[0032] Figure 3 It is the mAP curve graph and loss curve graph of the verification results of the present invention;

[0033] Figure 4 It is a F1score curve diagram of the verification result of the present invention. DETAILED DESCRIPTION

[0034] The following examples will enable those skilled in the art to more fully understand the present invention, but are not intended to limit the present invention in any way.

[0035] The present invention discloses a vibration signal positioning and identification method for a Φ-OTDR system, and the vibration signal positioning and identification method comprises the following steps:

[0036] Step 1: Use the Φ-OTDR system to collect external vibration signals, and mark the collected vibration signals with their categories and occurrence locations;

[0037] Step 2: Divide the labeled vibration signal data into fixed-size data and construct a space-time matrix sample data set, which includes a validation set and a training set.

[0038] Step 3, construct a fiber optic vibration signal positioning and classification model based on the sample size; the fiber optic vibration signal positioning and classification model includes four parts: a backbone network, a feature extraction and attention fusion network, a detection network and a post-processing structure; the backbone network uses multiple group convolutions to superimpose, and connects multiple residual modules in series after the group convolution to perform preliminary feature extraction and compression on the input vibration signal, and obtain multi-layer feature maps with different compression degrees, and output them to the feature extraction and attention fusion network; the feature extraction and attention fusion network uses the FPN structure to further extract features from the multi-layer feature maps, and combines the bottom-level features extracted by the backbone network with the FPN structure. The feature maps of different levels extracted by the N structure are fused, and a one-dimensional multi-layer compressed feature map is output as a predicted feature map and sent to the detection network; the detection network includes a first detection head for predicting the positioning adjustment parameters and a second detection head for predicting the classification probability, which respectively predict the positioning adjustment parameters and the classification probability, and the number of each detection head is the same as the number of layers of the multi-level feature map; the post-processing structure applies the positioning adjustment parameters to the preset length to obtain the signal position, filters the prediction results with low probability, sets a confidence threshold, and filters out the detection frames below the threshold. The threshold is usually set to 0.75 in field applications, and the detection frames below 0.75 will be discarded. The prediction results that exceed the input sample size are cropped, and the prediction results of the same category are processed by non-maximum suppression (NMS) to eliminate redundant detection frames.

[0039] Step 4: Use the training set to train the optical fiber vibration signal positioning and classification model. After the training is completed, use the validation set for verification and optimize the model hyperparameters based on the verification results.

[0040] The Φ-OTDR system of the present invention senses external vibration based on acoustic wave phase sensing. The laser emitted by a narrow line width distributed feedback fiber laser is modulated into pulse light by an acousto-optic modulator, and then amplified by an erbium-doped fiber amplifier and enters the sensing fiber. The backward Rayleigh scattering signal of the fiber enters the coupler through a circulator. Two ports on the other side of the coupler are connected to two Faraday rotator mirrors to form a Michelson interferometer. The incident light is divided into two beams and reflected by the Faraday rotator mirrors to interfere in the coupler. The interference signal is received by a detector and finally collected to a host computer for data processing. The host computer is used to collect and save vibration signals in real time, and the vibration signal positioning and identification method is used to realize the positioning and identification of external vibration signals.

[0041] Taking the application of pipeline external break alarm detection as an example, the vibration signal positioning and identification method of this example includes the following steps:

[0042] Step 1: You need to connect the Φ-OTDR system to the sensing optical cable of the monitoring pipeline and collect a certain number of vibration signals near the pipeline.

[0043] Step 2: Split the labeled data into the same size and construct a space-time matrix sample data set, which includes a validation set and a training set. The time length of the space-time matrix sample should try to include the complete vibration signal or reflect the characteristics of the vibration signal changing over time. Similarly, the spatial length of the space-time matrix sample should reflect the characteristics of the vibration signal changing over space.

[0044] Step 3: The overall network architecture is shown in Figure 1 ,According to the sample size, a fiber optic vibration signal positioning and classification model,was constructed. The model mainly includes four parts: backbone network, feature extraction and attention fusion network, detection network and post-processing.

[0045] The backbone network uses multiple group convolutions, followed by multiple residual modules in series to extract 4-level feature maps. In the backbone network, due to prior conditions, we divide the input into n groups (P1, P2, P3... Pi... Pn). Next, n convolution kernels (C1, C2, C3... Ci... Cn) are used to extract features from each group, and Ci only convolves the i-th part of the input group. The group convolution stage is described as follows:

[0046]

[0047] Where O is the output, and f is the nonlinear activation function that introduces nonlinear factors, which can increase the nonlinear fitting ability of the model. Due to the imbalance between the time dimension and the spatial dimension of the vibration data, G-BLOCK is used for initial feature extraction. This method can learn the vibration characteristics of different vibration positions and realize feature extraction in the time dimension while maintaining the invariance of the spatial dimension. After the group convolution, the original unbalanced input becomes a balanced feature map of size (100, 100). In order to improve the correlation between the outputs of the group convolution, the predicted feature maps of different receptive fields are obtained, and the FP-Block stacked by 4 RES-Blocks is added to generate the original feature pyramid. The sizes of the feature maps used to predict signals with different vibration ranges are different. Signals with a larger vibration range are predicted by a small feature map with a large receptive field, while signals with a narrower vibration range are predicted by a large feature map with a smaller receptive field.

[0048] The feature extraction and attention fusion network first uses the FPN (Feature Pyramid Networks) structure to extract feature maps at different levels, and then adds a fusion module based on the attention mechanism to further fuse the features. In order to improve the accuracy of vibration signal classification and positioning, feature fusion is performed between feature pyramids, and a feature attention fusion block (Fusion-A) is added to further strengthen the focus on the vibration signal positioning task. The input of the fusion module based on the attention mechanism is the lowest level feature extracted by the backbone network and the feature maps of different levels of the FPN structure. The number of fusion modules based on the attention mechanism is the same as the number of layers of the multi-level feature map. For the predicted feature map, the semantic information of the bottom-level features is relatively scarce, but the target location is accurate. The semantic information of the high-level features is rich, but the target location is rough. By combining the features of these different layers, the target location and high semantic information are fused together. In this process, the feature map of each layer is processed by 2DConv, and then the feature map of the upper layer is upsampled and added to the similar feature map of the lower layer. In addition, in order to effectively reduce the amount of calculation, we added g-conv in feature extraction and compressed the feature map. All output feature maps are compressed in the time dimension, and each 2D feature map is converted into a 1D feature map. Finally, in the Fusion-A module, all compressed feature maps are fused with the original 2D feature map of the lowest layer feature map to enhance the localization effect of the signal. The Fusion-A module contains g-conv for compressing the feature map of the original feature map and two 1×1 convolutions for adjusting the number of channels. The predicted feature map is obtained by multiplying the two output points.

[0049] The detection network consists of a detection head that predicts the positioning adjustment parameters and a detection head that predicts the classification probability, and the number of each detection head is the same as the number of layers of the multi-level feature map. In the detection network, the network generates the adjustment parameters and predicted probabilities of the anchor box for each predicted point. Both tasks are completed by two 1D convs of size 3. The output channels for classification and positioning are 6 and 2, respectively, representing the number of object categories and the adjustment parameters. Each point of the predicted feature map is used to predict on an anchor box with a default length of three lengths. First, in order to simplify the task of position prediction, a set of default lengths is established at each predicted point in the predicted feature map. The default length depends on the receptive field of the feature map.

[0050] The detection network consists of a detection head that predicts positioning adjustment parameters and a detection head that predicts classification probability. The number of each detection head is the same as the number of layers in the multi-level feature map.

[0051] Post-processing mainly performs four tasks: applying prediction adjustment parameters to the default length; filtering low-probability prediction results; cropping prediction results that exceed the input sample size, and performing NMS processing on prediction results of the same category.

[0052] During the model training process, loss functions are set for positioning and classification tasks respectively. The total loss function is the linear superposition of the positioning loss function and the classification task loss function. Stochastic gradient descent method is used for iterative training.

[0053] The final network architecture details are as follows Figure 2 exhibit.

[0054] Step 4: Use the training data set to train the optical fiber vibration signal positioning and classification model. After the training is completed, use the validation set for verification. According to the actual situation of the data set, the model is finally selected. In the prediction process, a maximum of 10 prediction targets are output from high to low according to the prediction probability. By comparing the predicted target with the real sample label, the number of correct matches TP (True Positive), the number of missed detections FN (False Negative) and the number of wrong (repeated) detections FP (False Positive) are calculated. The two factors that affect the matching are the prediction probability threshold and the IOU (Intersection over Union) threshold of the predicted length and the real length. That is, assuming that the predicted category is correct, the match is successful only when it exceeds a certain prediction probability threshold and the IOU of the predicted length and the real length also exceeds a certain threshold, that is, TP. Then, the precision rate and recall rate can be calculated. In the task of monitoring pipeline leakage, they are related to the false alarm rate NAR (Nuisance Alarm Rate) and the missed alarm rate MAR (Missing Alarm Rate), respectively.

[0055]

[0056]

[0057] In practical applications, the model must consider both the false alarm rate and the missed alarm rate. Therefore, the F1score can be obtained by comprehensively considering the precision and recall.

[0058]

[0059] Similarly, using precision and recall as the horizontal and vertical coordinates to obtain the PR curve, the area under the curve is AP (Average Precision), and the average AP of all categories is map (Mean Average Precision). Assume that the IOU threshold for successful matching is set to 0.5, and the prediction probability threshold starts from 0 and ends at 1, with an interval of 0.01 for a total of 100 values. During the entire training process, we use the validation set to verify and evaluate the trained network after each round of training iteration to obtain the mAP under different conditions, and then sum and average the obtained mAP at 100 probability thresholds. Finally, we get the training loss and the averaged mAP curve. As shown Figure 3 shown.

[0060] It can be seen from the curve that the map can reach up to 99.3%, reaching a relatively high value, indicating that the proposed method has high precision and recall rates, while the false alarm rate and missed alarm rate are very low.

[0061] Furthermore, the preset TP condition is IOU = 0.5, and the F1scores under different prediction probability thresholds are plotted, such as Figure 4 As shown in the figure, we can see that the highest F1score is 99% when the prediction probability is 0.621. Therefore, when post-processing the model, we can filter out the prediction targets below 0.621, which will reduce the need for NMS operations, speed up the prediction speed, and improve the overall prediction effect of the method.

[0062] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A vibration signal positioning and identification method for a Φ-OTDR system, characterized in that: The vibration signal positioning and identification method comprises the following steps: Step 1: Use the Φ-OTDR system to collect external vibration signals, and mark the collected vibration signals with their categories and occurrence locations; Step 2: Split the labeled vibration signal data into fixed-size data and construct a space-time matrix sample data set. The space-time matrix sample data set includes a validation set and a training set. All samples in the data set are space-time matrices of the same size, and each sample has a category label and a vibration position label. Step 3, constructing an optical fiber vibration signal positioning and classification model according to the sample size; the optical fiber vibration signal positioning and classification model includes four parts: a backbone network, a feature extraction and attention fusion network, a detection network and a post-processing structure; the backbone network uses multiple group convolutions to superimpose, and connects multiple residual modules in series after the group convolution to perform preliminary feature extraction and compression on the input vibration signal, obtain multi-layer feature maps with different compression degrees, and output them to the feature extraction and attention fusion network; the feature extraction and attention fusion network uses the FPN structure to further extract features from the multi-layer feature map, and fuses the bottom-level features extracted by the backbone network and the feature maps of different levels extracted by the FPN structure, outputs a one-dimensional multi-layer compressed feature map as a predicted feature map and sends it to the detection network; the detection network includes a first detection head for predicting positioning adjustment parameters and a second detection head for predicting classification probability, which predict positioning adjustment parameters and classification probability respectively, and the number of each detection head is the same as the number of layers of the multi-level feature map; the post-processing structure applies the positioning adjustment parameters to the preset length to obtain the signal position, filters the low-probability prediction results, clips the prediction results that exceed the input sample size, and performs non-maximum suppression processing on the prediction results of the same category; Step 4: Use the training set to train the optical fiber vibration signal positioning and classification model. After the training is completed, use the validation set for verification and optimize the model hyperparameters based on the verification results.

2. The vibration signal positioning and identification method for a Φ-OTDR system according to claim 1, characterized in that: In step 1, the categories of the vibration signal include at least five categories: touching the optical fiber, walking, manual digging, excavator digging and climbing.

3. The vibration signal positioning and identification method for a Φ-OTDR system according to claim 1, characterized in that: In step 2, the backbone network uses multiple group convolution blocks G-BLOCK to perform preliminary feature extraction and compression on the input vibration signal, and the process of obtaining multi-layer feature maps with different compression degrees includes the following steps: The input vibration signal is divided into n groups, and the convolution kernels of n group convolution blocks G-BLOCK extract features from each group. The convolution kernel of each group convolution block G-BLOCK performs corresponding convolution on only one of the input groups to learn the vibration characteristics of different vibration positions and extract features from the time dimension while maintaining the invariance of the spatial dimension; the group convolution stage is described as follows: Where O is the output, f is a nonlinear activation function that introduces a nonlinear factor to increase the nonlinear fitting ability of the model, C, w_i, b_i and x_Parti represent the number of convolution kernels, convolution weights, convolution bias and the input of the i-th group respectively; An FP-Block consisting of four RES-Blocks stacked together is added to generate the original feature pyramid to improve the correlation between the outputs of the group convolution and obtain predicted feature maps with different receptive fields; among them, the signal with a larger vibration range is predicted by a small feature map with a large receptive field, and the signal with a narrower vibration range is predicted by a large feature map with a smaller receptive field.

4. The vibration signal positioning and identification method for a Φ-OTDR system according to claim 1, characterized in that: In step 2, the process of outputting a one-dimensional multi-layer compressed feature map from the feature extraction and attention fusion network as a prediction feature map and sending it to the detection network includes the following steps: The feature maps of each layer are processed by 2DConv, and then the feature maps of the upper layer are upsampled and added to the similar feature maps of the lower layer, so that the target position and high semantic information are fused together; Add multiple g-conv layers, each of which compresses each fused feature map, compresses all output feature maps in the time dimension, and converts each two-dimensional feature map into a one-dimensional feature map; In the Fusion-A module, all compressed one-dimensional feature maps are fused with the original 2D feature map of the lowest layer feature map to enhance the signal positioning effect; the Fusion-A module contains a g-conv layer for compressing the original feature map and two 1×1 convolution layers for adjusting the number of channels. A one-dimensional multi-layer compressed feature map is obtained by dot multiplication of the outputs of the two 1×1 convolution layers, which is output to the detection network as a predicted feature map.

5. The vibration signal positioning and identification method for a Φ-OTDR system according to claim 1, characterized in that: In step 2, the detection network includes a first detection head for predicting the positioning adjustment parameter and a second detection head for predicting the classification probability. The process of respectively predicting the positioning adjustment parameter and the classification probability includes the following steps: A set of default length values ​​is established at each prediction point of the predicted feature map, and the default length value depends on the receptive field of the feature map; each point of the predicted feature map is used to make predictions on anchor boxes with three default lengths; Two 1D convs of size 3 are used to generate the adjustment parameters and prediction probabilities of the anchor boxes for each predicted point. The number of output channels of the 1D convs used for classification and positioning is the number of object categories and the number of positioning adjustment parameters, respectively.

6. The vibration signal positioning and identification method for a Φ-OTDR system according to claim 1, characterized in that: In step 4, during the training process of the optical fiber vibration signal positioning and classification model, loss functions are set for the positioning and classification tasks respectively. The total loss function is the linear superposition of the positioning task loss function and the classification task loss function. The random gradient descent method is used to iteratively train the optical fiber vibration signal positioning and classification model.

Images