Semi-supervised DAS event detection method and system

Abstract

The invention relates to the technical field of distributed optical fiber sensing, and discloses a semi-supervised DAS event detection method and system, and the method comprises the following steps: selecting a data set collected by a DAS system, and carrying out the preprocessing of the data set, and obtaining paired training data; building a semi-supervised DAS event detection model which comprises a core event detection network, a student network and a teacher network; designing a loss function and carrying out learning training on the semi-supervised DAS event detection network; and inputting a new group of to-be-detected data into the trained event detection network to obtain an event classification and positioning result so as to realize event detection. According to the semi-supervised architecture of the designed DAS event detection model, two semi-supervised technologies of pseudo-tag and consistency regularization are combined, a large amount of unlabeled DAS data and a small amount of labeled DAS data are fully utilized for joint training, the dependence on the labeled DAS data is reduced, the labeling cost is remarkably reduced, and meanwhile the model performance is improved.

Description

Technical Field

[0001] The present invention relates to the field of distributed optical fiber sensing technology, and in particular to a semi-supervised DAS event detection method and system. Background Art

[0002] Distributed fiber acoustic sensing (DAS) technology based on phase-sensitive optical time-domain reflectometry (OTDR) has the advantages of being able to provide long-distance, high-resolution acoustic signal monitoring, continuous large-range detection, and positioning of acoustic signals acting on the sensing optical fiber, high sensitivity, and fast response. It has attracted close attention and has been widely used.

[0003] By analyzing the changes in the optical signal propagating in the optical fiber, the DAS system can detect and locate events along the optical fiber path, such as intrusion, vibration or temperature changes. Using the widely laid underground communication cables, various vibration sources can be sensed and located with high sensitivity and high accuracy. Therefore, the DAS system has been applied in various underground and ground monitoring, such as natural disaster prediction of submarine seismic activity, volcanic events and earthquakes, energy exploration in the oil and gas industry, and civil infrastructure safety monitoring in energy transmission pipelines, railways, perimeter security, and safety monitoring of urban underground structures.

[0004] In recent years, with the continuous development of artificial intelligence technology, security threat event detection technology based on deep learning has also emerged. A large number of studies have shown that models built based on deep learning methods can achieve good performance in security threat event detection problems. However, in practical applications, due to the wide monitoring range and high resolution of DAS, it is very time-consuming to obtain a large amount of labeled DAS data. It is often impractical to obtain enough labeled samples for training deep learning models in practical applications. Semi-supervised learning can use a large amount of unlabeled data to improve the performance of the model. Semi-supervised learning is a machine learning strategy that combines the advantages of supervised learning and unsupervised learning. It reduces the cost of data annotation by using a small amount of labeled data and a large amount of unlabeled data, while improving the generalization ability and prediction accuracy of the model. This method is particularly suitable for scenarios where it is costly or time-consuming to obtain a large amount of labeled data, such as medical diagnosis, natural language processing, and image recognition. The advantage of semi-supervised learning is that it can use data more effectively, reduce dependence on labeled data, improve data usage efficiency, enhance the robustness of the model, and maintain the scalability of model performance in the big data era. In this way, the model can better understand and predict the inherent structure and relationships of the data, thereby providing more accurate and reliable results in real-world applications. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a semi-supervised DAS event detection method and system.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] A semi-supervised DAS event detection method, wherein the DAS event detection model adopted includes a core event detection network, and a student network and a teacher network formed by the core event detection network; the core event detection network includes a backbone network, a neck network and a detection head;

[0008] The student network and the teacher network form a semi-supervised architecture for training: the student network is trained using labeled DAS data, the learnable weights are updated through back-propagation, and the teacher network is gradually updated by applying exponential moving average; the teacher network predicts the weakly enhanced unlabeled DAS data, and the prediction results above the set event classification confidence threshold are added to the training as DAS event classification pseudo-labels, so as to train the DAS event classification branch on the unlabeled DAS data, and iteratively update the weights of the student network; the prediction results below the set interval regression variance threshold are used as DAS event pseudo-interval labels, so as to train the interval DAS event regression branch on the unlabeled DAS data, and iteratively update the weights of the student network; the student network is optimized through supervision loss, classification loss of DAS event classification pseudo-labels, and regression loss of DAS event pseudo-interval labels, and the parameters of the teacher network are only updated by the student network through exponential moving average;

[0009] The DAS data is passed through the backbone network for feature extraction in turn, and the obtained feature map is upsampled and fused through the neck network to obtain a preliminary prediction result. Finally, the final DAS event classification prediction result and DAS event positioning prediction result are obtained through the DAS event classification branch and DAS event regression branch of the detection head respectively.

[0010] Furthermore, by building a distributed optical fiber sensing data acquisition platform, the monitoring signal of the DAS event is acquired to obtain the DAS data.

[0011] Furthermore, the DAS data is sequentially subjected to feature extraction through the backbone network, specifically including:

[0012] The DAS data undergoes downsampling operation N1 to output feature map group C1, C1 undergoes multiple feature fusion downsampling operations N2 in turn to output feature map groups C2, C3, and C4, and C4 undergoes convolutional attention fusion operation N3 to output feature map group C5.

[0013] Furthermore, the downsampling operation N1 is composed of a one-dimensional convolutional layer Conv1d, a batch normalization BatchNorm1d and an activation function SiLU; the feature fusion downsampling operation N2 is composed of a downsampling operation N1 and a segmentation operation Split, a concatenation operation Concat, and an addition operation Add; the convolutional attention fusion operation N3 is composed of a downsampling operation N1, a maximum pooling layer Maxpooling, an attention module Attention, and a feedforward neural network FFN.

[0014] Furthermore, the feature map is obtained by upsampling the neck network and fusing the features to obtain a preliminary prediction result, which specifically includes:

[0015] After the upsampling operation N4, the feature map group C5 is merged with the feature map group C4 by channel, and the feature fusion N5 outputs the feature map group F4. After the upsampling operation N4, F4 is merged with the feature map group C3 by channel, and the preliminary prediction result P3 is output through N5. After the downsampling operation N1, P3 is merged with F4 by channel, and the preliminary prediction result P4 is output through N5. After the downsampling operation N1, P4 is merged with C5 by channel, and the preliminary prediction result P5 is output through feature fusion N5.

[0016] Furthermore, the feature fusion operation N5 is composed of a one-dimensional convolution layer, batch normalization and SiLU activation function, as well as a segmentation operation, a splicing operation, and an addition operation.

[0017] Furthermore, the final event classification prediction result and event location prediction result obtained by the detection head specifically include:

[0018] The preliminary prediction results P3, P4, and P5 are respectively subjected to multi-dimensional convolution and compound loss optimization operation N6 to obtain the final event classification prediction results and event location prediction results; the multi-dimensional convolution and compound loss optimization operation N6 consists of downsampling operation N1, depthwise separable convolution DWConv, and IOU loss, distribution focus loss, and binary cross entropy loss BCE.

[0019] Furthermore, the student network is optimized by supervision loss, classification loss of DAS event classification pseudo labels and regression loss of DAS event pseudo interval labels, and the parameters of the teacher network are only updated by the student network through exponential moving average, specifically including:

[0020] The labeled DAS data is weakly enhanced and then input into the student network for supervised training, and the supervision loss L is calculated. s ;

[0021]

[0022] in, represents the θth labeled DAS data, is the classification loss of the DAS event classification label, is the regression loss of the DAS event interval label;

[0023] The unlabeled DAS data are processed by strong enhancement and weak enhancement respectively. The weakly enhanced DAS data are input into the teacher network for prediction. The prediction results higher than the set event classification confidence threshold are used as DAS event classification pseudo labels, and the classification loss of the unlabeled DAS data is calculated.

[0024]

[0025] in, The i-th unlabeled DAS data, express The corresponding DAS event pseudo labels generated by the teacher network for classification;

[0026] For the prediction results below the set interval regression variance threshold, they are used as pseudo interval labels for DAS events, and the regression loss of unlabeled DAS data is calculated.

[0027]

[0028] in express The corresponding pseudo interval labels of DAS events generated by the teacher network for regression;

[0029] Finally, the total unsupervised loss is obtained

[0030]

[0031] In each iteration of the training phase, the student network passes through L s and L u The total loss L composed of weighted fusion is optimized:

[0032]

[0033] λ u Parameters contributing to controlling the unsupervised loss;

[0034] The parameters of the teacher network are updated by the student network through exponential moving average, with weights θ s Student Model M s and weight θ t Teacher Model M t ,θ s and θ tThe exponential moving average update process between is as follows:

[0035]

[0036] where the superscript i denotes the i-th iteration during training and α denotes the smoothing coefficient of the exponential moving average.

[0037] A semi-supervised DAS event detection system, wherein the DAS event detection model adopted includes a core event detection network, and a student network and a teacher network formed by the core event detection network; the core event detection network includes a backbone network, a neck network and a detection head; the semi-supervised DAS event detection system specifically includes:

[0038] In the training module, the student network and the teacher network form a semi-supervised architecture for training: the student network is trained using labeled DAS data, the learnable weights are updated through back-propagation, and the teacher network is gradually updated by applying exponential moving average; the teacher network predicts the weakly enhanced unlabeled DAS data, and the predicted results above the set event classification confidence threshold are added to the training as DAS event classification pseudo-labels, so as to train the DAS event classification branch on the unlabeled DAS data, and iteratively update the weights of the student network; the predicted results below the set interval regression variance threshold are used as DAS event pseudo-interval labels, so as to train the interval DAS event regression branch on the unlabeled DAS data, and iteratively update the weights of the student network; the student network is optimized through the supervision loss, the classification loss of the DAS event classification pseudo-labels, and the regression loss of the DAS event pseudo-interval labels, and the parameters of the teacher network are only updated by the student network through exponential moving average;

[0039] In the prediction module, the DAS data is extracted through the backbone network in sequence, and the obtained feature map is upsampled and fused through the neck network to obtain a preliminary prediction result. Finally, the final DAS event classification prediction result and DAS event positioning prediction result are obtained through the DAS event classification branch and DAS event regression branch of the detection head respectively.

[0040] The system of the present invention corresponds to the method, and the specific technical solution of the method is also applicable to the system.

[0041] Compared with the prior art, the beneficial technical effects of the present invention are:

[0042] The semi-supervised architecture of the DAS event detection model designed by the present invention combines two semi-supervised technologies, pseudo-labeling and consistency regularization, to make full use of a large amount of unlabeled DAS data and a small amount of labeled DAS data for joint training, reduce the dependence on labeled DAS data, significantly reduce the labeling cost, and improve the model performance at the same time. In the pseudo-label part, the student network is used to predict the unlabeled data, generate pseudo-labels, and only the pseudo-labels with high prediction confidence are input into the teacher network through screening. Through the combination of strong enhancement and weak enhancement, the model's prediction consistency for the same data sample is maintained, so that the model can maximize the potential of unlabeled data and further improve the classification accuracy and detection accuracy of the model. In the consistency regularization part, the unlabeled data below the confidence threshold is input into the teacher network, forcing its prediction results to be close to the prediction of the teacher network, reducing the risk of overfitting, and improving the robustness, stability and generalization ability of the model. By using the average prediction results of the teacher network, the prediction oscillation of the model on unlabeled samples can be reduced, thereby stabilizing the training process. The detection accuracy is improved by the weighted averaging method, and the detection performance of the model is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is the overall network framework diagram of the present invention.

[0044] Figure 2 It is a schematic diagram of the DAS event detection model in the present invention.

[0045] Figure 3 Schematic diagram of the semi-supervised architecture (student network and teacher network) in the present invention.

[0046] Figure 4 It is a flow chart of the downsampling operation N1, the downsampling operation N2, and the convolutional attention fusion operation N3 in the present invention; wherein, (a) is the downsampling operation N1, (b) is the downsampling operation N2, and (c) is the convolutional attention fusion operation N3.

[0047] Figure 5 It is a flow chart of the feature fusion operation N5, multi-dimensional convolution and compound loss optimization operation N6 in the present invention; wherein, (a) is the feature fusion operation N5, and (b) is the multi-dimensional convolution and compound loss optimization operation N6.

[0048] Figure 6 Schematic diagram of DAS data of five event types in an embodiment of the present invention; wherein (a) is DAS data of a manual hoeing event type, (b) is DAS data of a manual drilling event type, (c) is DAS data of an excavator working event type, (d) is DAS data of a road breaker working event type, and (e) is DAS data of a vehicle driving event type.

[0049] Figure 7 It is the confusion matrix corresponding to each model in the data set of Table 2; Among them, (a) is the confusion matrix corresponding to the FPN-BiLSTM model, (b) is the confusion matrix corresponding to the CNN-BiGRU model, (c) is the confusion matrix corresponding to the STF-XGB model, (d) is the confusion matrix corresponding to the DSCDN model, (e) is the confusion matrix corresponding to the MTL model, and (f) is the confusion matrix corresponding to the DAS event detection model of the present invention.

[0050] Figure 8 This is a schematic diagram of the performance comparison of different semi-supervised methods when the label amount is 1%, 2%, 5%, and 10%.

[0051] Fig. 9 2 is a schematic diagram of the detection results of DAS data of different event types according to the present invention; wherein (a) is the detection result of DAS data of manual hoeing event type, (b) is the detection result of DAS data of manual drilling event type, (c) is the detection result of DAS data of excavator working event type, (d) is the detection result of DAS data of road breaker working event type, and (e) is the detection result of DAS data of vehicle driving event type.

[0052] Fig.10 It is a schematic diagram of the effects of using different hyperparameters in the present invention. DETAILED DESCRIPTION

[0053] A preferred embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0054] Example 1

[0055] The overall process of the present invention is as follows Figure 1 This embodiment discloses a semi-supervised DAS event detection method, including steps S1 to S4, which are as follows:

[0056] S1. Collect DAS data and preprocess them to obtain training data sets.

[0057] Step S1 specifically includes the following steps:

[0058] S11. Build a distributed optical fiber sensing data acquisition platform, including obtaining monitoring signals of mechanical excavation, manual hoeing, vehicle driving and other events, and recording original DAS data;

[0059] S12. Perform preprocessing such as denoising, labeling, weak enhancement, and strong enhancement on the collected original DAS data to obtain a training data set and a test data set.

[0060] S2. Build a semi-supervised DAS event detection model, including the core event detection network, student network and teacher network.

[0061] Step S2 specifically includes the following steps:

[0062] S21. Design the core event detection network (Event Classification and Localization Networks, ECLNet):

[0063] like Figure 2 As shown in the figure, for a given input DAS data X, the feature map is obtained by extracting features through the backbone network Backbone, and then the preliminary prediction result is obtained through upsampling and feature fusion of the neck network Neck. Finally, the final classification prediction result and positioning prediction result are obtained through the detection head Head.

[0064] The backbone network Backbone part is as follows: the original DAS data X is downsampled by N1 to output the feature map group C1, C1 is downsampled by multiple N2 in turn to output the feature map groups C2, C3, and C4, and C4 is downsampled by N3 to output the feature map group C5.

[0065] Among them, N1 consists of a one-dimensional convolution layer Conv1d, batch normalization BatchNorm1d and activation function SiLU. The one-dimensional convolution kernel size is 5×1, the step size is 2, and the padding is automatically calculated. N2 consists of a downsampling operation N1, a segmentation operation Split, a concatenation operation Concat, and an addition operation Add. N3 consists of a downsampling operation N1, a maximum pooling layer Maxpooling, an attention module Attention, and a feedforward neural network FFN.

[0066] For the specific structures of N1, N2 and N3, please refer to Figure 4 (a), (b), and (c) in the figure.

[0067] The overall Neck part of the neck network is as follows: C5 is upsampled by N4 and merged with C4 by channel, and then N5 outputs the feature map group F4. F4 is upsampled by N4 and merged with C3 by channel, and then N5 outputs the preliminary prediction result P3. P3 is downsampled by N1 and merged with F4 by channel, and then N5 outputs the preliminary prediction result P4. P4 is downsampled by N1 and merged with C5 by channel, and then N5 outputs the preliminary prediction result P5.

[0068] N4 is an upsampling module. Figure 5 As shown in (a) in the figure, N5 consists of a one-dimensional convolutional layer, batch normalization and SiLU activation function, as well as operations such as Split, Concat, and Add.

[0069] The structure of the detection head is as follows:

[0070] The preliminary prediction results P3, P4, and P5 are respectively passed through N6 to obtain the final classification results and positioning results. The convolution of classification and positioning is decoupled. One convolution branch is the positioning regression branch. The number of channels of the feature map in the positioning interval is 2, that is, one event start position and one event end position; the number of channels of classification is the number of categories. Classification uses depthwise separable convolution instead of conventional convolution to reduce the amount of calculation.

[0071] The preliminary prediction results P3, P4, and P5 are also feature maps, but they are more abstract after more upsampling and downsampling operations.

[0072] like Figure 5 As shown in (b), N6 consists of the downsampling operation N1, the depth-wise separable convolution DWConv, as well as the IOU loss, the distribution focus loss, and the binary cross entropy loss BCE.

[0073] The IOU is:

[0074]

[0075] Among them, I pre is the prediction interval, I gt is the true interval.

[0076]

[0077] Among them, y i is the true label, p i is the predicted probability. DFL is obtained by adjusting p i The weight of makes the model pay more attention to samples that are difficult to classify.

[0078]

[0079] Among them, y i is the true label of sample i (0 or 1), p i is the probability that sample i is predicted to be positive, and N is the number of samples.

[0080] S22. Define the semi-supervised architecture (student network and teacher network):

[0081] First, the input DAS data is preprocessed to obtain a labeled data set and an unlabeled data set. For each DAS data in the labeled data set, basic weak data enhancement is performed to increase the robustness of the model. For each DAS data in the unlabeled data set, two data enhancement methods with different strengths are used: one is weak enhancement and the other is strong enhancement. The main structure of the teacher network and the student network is consistent, and both are composed of a core event detection network. The trained student network model is used to classify and locate each DAS data in the unlabeled data set, and its classification probability is calculated. If the maximum probability value of the predicted classification result of a certain unlabeled DAS data exceeds the preset threshold, the predicted result is used as a pseudo label and added to the training data set. Specifically, the pseudo label generation rule is: set a minimum probability threshold τ, and only when the student network predicted classification probability is greater than the threshold, the predicted result is used as a valid pseudo label. After that, the DAS data with real labels and the DAS data with pseudo labels are mixed, and the weight of the teacher network is iteratively updated to improve the performance of the model in data classification accuracy. This part of the loss calculates the error between the DAS data in the unlabeled dataset and the generated pseudo-labels, using the same cross-entropy loss function as the classification loss for labeled data.

[0082] For DAS data with prediction probabilities lower than the threshold, according to the consistency regularization principle, the model is required to generate consistent prediction results for the same unlabeled DAS data on different transformed versions after weak enhancement and strong enhancement. That is, for each unlabeled DAS data, the model's prediction should be consistent with its weak enhancement version and strong enhancement version. If the prediction results are inconsistent, a penalty term is added to make the model more robust to the transformation of the input data. In this process, the outputs of the teacher network and the student network are compared through consistency loss to ensure that the student network can obtain accurate guidance information from the teacher network during the learning process. Specifically, by calculating the difference between the outputs of the teacher and student networks and minimizing it, the student network can gradually approach the output of the teacher network.

[0083] S3. Define the loss function and perform network training.

[0084] Step S3 specifically includes the following steps:

[0085] S31. Define supervision loss L lab :

[0086] Supervision loss L labThe calculation includes two branches: classification and regression. The classification branch uses BCE Loss. The regression branch uses Distribution Focal Loss (DFL) loss and DIoU loss. The three losses are weighted and summed with a certain weight ratio to get the supervision loss.

[0087] S32. Define training strategy:

[0088] In the training phase, the labeled DAS data is weakly enhanced and then input into the student network for supervised training, and the supervision loss L is calculated. s ;

[0089]

[0090] in, represents the i-th labeled DAS data, is the classification loss of the DAS event classification label, is the regression loss of the DAS event interval label;

[0091] The unlabeled DAS data are processed by strong enhancement and weak enhancement respectively. The weakly enhanced DAS data are input into the teacher network for prediction. The prediction results higher than the set event classification confidence threshold are used as DAS event classification pseudo labels, and the classification loss of the unlabeled DAS data is calculated.

[0092]

[0093] in, The i-th unlabeled DAS data, express The corresponding DAS event pseudo labels generated by the teacher network for classification;

[0094] For the prediction results below the set interval regression variance threshold, they are used as pseudo interval labels for DAS events, and the regression loss of unlabeled DAS data is calculated.

[0095]

[0096] in express The corresponding pseudo interval labels of DAS events generated by the teacher network for regression;

[0097] Finally, the total unsupervised loss is obtained

[0098]

[0099] In each iteration of the training phase, the student network passes through L s and L u The total loss L composed of weighted fusion is optimized:

[0100]

[0101] λ u Parameters contributing to controlling the unsupervised loss;

[0102] The parameters of the teacher network are updated by the student network through the exponential moving average (EMA). The present invention updates the EMA between the student model Ms with a weight of θs and the teacher model Mt with a weight of θt as follows:

[0103]

[0104] Where superscript i denotes the i-th iteration during training and α denotes the smoothing coefficient of EMA.

[0105] S4. Model application:

[0106] After the training phase is completed, the DAS data in the test dataset is directly processed through the core event detection network, producing the final event classification and event localization results without additional processing steps.

[0107] Example 2

[0108] The present invention also discloses a semi-supervised DAS event detection system. The DAS event detection model used includes a core event detection network, and a student network and a teacher network formed by the core event detection network; the core event detection network includes a backbone network, a neck network and a detection head; the semi-supervised DAS event detection system specifically includes:

[0109] The training module is trained by forming a semi-supervised architecture with a student network and a teacher network. The student network predicts the unlabeled DAS data, and the event classification prediction results with a prediction probability higher than the set value are added to the training as pseudo-labels, and the weights of the teacher network are iteratively updated. For DAS data with a prediction probability lower than the set value, consistency regularization is used, requiring the student network and the teacher network to have consistent prediction results for the weakly enhanced and strongly enhanced versions of the unlabeled DAS data, respectively. In the training phase, the student network and the teacher network are optimized through supervision loss, pseudo-label cross entropy loss, and consistency loss.

[0110] In the prediction module, the DAS data is extracted through the backbone network in sequence, and the obtained feature map is upsampled and fused through the neck network to obtain a preliminary prediction result. Finally, the final event classification prediction result and event location prediction result are obtained through the detection head.

[0111] The system of the present invention corresponds to the method, and the specific technical solution of the method is also applicable to the system.

[0112] Experimental results:

[0113] The event signal samples collected on site are as follows: Figure 6 As shown, Figure 6 (a), (b), (c), (d), and (e) are five events: manual hoeing, manual drilling, excavator work, road breaker work, and vehicle driving. The database used for the experiment is shown in Table 1.

[0114] Table 1. Database constructed for five types of target event recognition:

[0115] Label Event Type Training set Validation set Test Set Total dataset 1 Manual hoeing 7221 1444 1444 10109 2 Artificial drilling 7193 1439 1439 10070 3 Excavator work 7441 1488 1488 10417 4 Road breaker work 7389 1478 1478 10345 5 Vehicle driving 7346 1469 1469 10284

[0116] The method of the present invention is compared with a variety of methods to analyze the contribution of the present invention. It is worth mentioning that, in theory, the semi-supervised framework of the present invention is suitable for tasks such as event classification and detection, but the current research on DAS data is limited to the classification of events, and few methods for realizing positioning tasks have been proposed. This work can be said to be the first real research on using DAS data for event detection. Therefore, in the performance comparison part of supervised learning, the present invention only compares the classification performance of the model's test core event detection network ECLGNet.

[0117] In Table 2, the performance of the proposed algorithm in terms of accuracy, precision, recall and F1 value is illustrated.

[0118]

[0119]

[0120]

[0121] Where FP, TP, FN and TN represent the number of false positive samples, true positive samples, false negative samples and true negative samples respectively.

[0122] The results show that the network introduced by the present invention shows excellent performance in various evaluation indicators in the field of multi-label event classification. It is worth noting that the proposed method outperforms other classification networks in accuracy, precision, recall and F1 value, reaching results of 96.31%, 96.33%, 96.32% and 96.33% respectively. These results emphasize the effectiveness of the method of the present invention specifically for multi-label event classification.

[0123] Table 2, performance comparison with other classification models:

[0124]

[0125]

[0126] To illustrate the performance of the model classification in more detail, a confusion matrix is ​​plotted on the constructed dataset, as shown in Figure 7 As shown. The method of the present invention is evaluated by comparing it with the above-mentioned state-of-the-art models. It is obvious that the method proposed in the present invention outperforms other state-of-the-art multi-label classification networks. Regarding the average accuracy of each event category, this paper achieves the best performance in the five event categories of manual hoeing, manual drilling, excavator work, road breaker work and vehicle driving, with average accuracy scores of 97.78%, 97.50%, 95.16%, 95.33% and 95.85% respectively. Among them, Figure 7 (a), (b), (d), (d), and (e) are the confusion matrices corresponding to the FPN-BiLSTM model, CNN-BiGRU model, STF-XGB model, DSCDN model, MTL model, and the DAS event detection model of the present invention, respectively.

[0127] Next, the proposed method is experimented with several current advanced semi-supervised methods on datasets with 1%, 2%, 5%, and 10% annotation labels, and the performance is compared using mAP as the evaluation indicator.

[0128]

[0129]

[0130] Where P is the precision, R is the recall, and N is the number of categories.

[0131] The results are shown in Table 3, revealing the effectiveness of semi-supervised methods in utilizing unlabeled data. The results are visualized as follows Figure 8As shown in Figure 2, the performance of the method of the present invention is better than all the comparison methods in all label quantities. It is worth mentioning that in order to make a fair comparison with other semi-supervised methods, the ECLNet designed by the present invention is uniformly used as the default detector framework to illustrate the method of the present invention. In addition, a method similar to the object detection task is used to visualize the results of the present invention, as shown in Figure 2. Fig. 9 As shown, in order to more intuitively feel and compare the performance of the method of the present invention. Fig. 9 (a), (b), (c), (d), and (e) represent manual hoeing, manual drilling, excavator operation, road breaker operation, and vehicle driving, respectively.

[0132] Table 3, performance comparison of 1%, 2%, 5%, 10% label amount and other semi-supervised models:

[0133]

[0134]

[0135] In order to analyze the effectiveness of the key components in the method designed by the present invention, the present invention conducted an ablation study to verify the key design of the present invention. All ablation experiments were performed on the DAS dataset with 10% labels provided by the present invention.

[0136] 1) Influence of each component: The present invention removes some components in the semi-supervised method respectively and observes the changes in model performance. The present invention eliminates the influence of exponential moving average, classification pseudo-label and pseudo-interval label. The results are shown in Table 4. By applying the classification pseudo-label strategy, the method of the present invention achieves 65.28mAP, and then the teacher model is updated with the student model through the exponential moving average strategy of the teacher-student model. The method of the present invention further improves by 14.85 points. Further application of pseudo-interval labels, the performance reaches 85.44mAP, which is 5.31 points better than classification pseudo-label plus exponential moving average.

[0137] Table 4, Ablation studies:

[0138]

[0139]

[0140] 2) Influence of hyperparameters: This section is mainly used to explore the influence of hyperparameters used in the method of the present invention. Fig.10 shown.

[0141] Fig.10(a) in the figure shows the impact of different event classification confidence thresholds. The best performance is achieved when the threshold is set to 0.9, and lower or higher thresholds will lead to a significant drop in performance. When the threshold varies from 0 to 0.9, as expected, the number of pseudo-labels generated increases as the threshold decreases. Models using too high a threshold (e.g., 0.95) cannot obtain satisfactory results because the number of pseudo-labels generated is very small. On the other hand, models using low thresholds (e.g., 0.7) also cannot obtain good results because the model generates too many pseudo-labels, which are likely to be false positive instances. The present invention also observed that if the threshold is lower than 0.5, the model cannot even converge.

[0142] like Fig.10 As shown in (b), the present invention studies the influence of the interval regression variance threshold. The best performance is achieved when the threshold is set to 0.02.

[0143] The present invention also studies the influence of EMA rate α. The present invention uses various EMA rates from 0.5 to 0.9999 to evaluate the model and Fig.10 (c) in the figure shows the mAP results of the teacher model. The present invention observes that with a smaller EMA rate (e.g., α = 0.5), the model has a lower mAP because the student contributes more to the teacher model in each iteration. This means that the teacher model may be adversely affected by noisy pseudo-labels. As α increases, this unstable learning curve can be stabilized and improved. When α reaches 0.999, it exhibits the best mAP. However, if α continues to increase, the teacher model will grow too slowly because the teacher model mainly derives the next model weights from the previous teacher model weights.

[0144] To check the unsupervised loss weight λ u The present invention changes the unsupervised loss weight from 1.0 to 8.0 in the case of 10% labeled data. Fig.10 As shown in (d) in Figure 3, the model performs poorly at lower unsupervised loss weights. On the other hand, the present invention observes that the model performs best when the unsupervised loss weight is 5.0. However, when the weight is increased to 8.0, the training of the model fails to converge.

[0145] Experimental results show that these components are crucial to improving model performance, and their fusion significantly improves the detection performance of the model.

[0146] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0147] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A semi-supervised DAS event detection method, characterized in that: The DAS event detection model used includes a core event detection network, and a student network and a teacher network composed of the core event detection network; the core event detection network includes a backbone network, a neck network and a detection head; The student network and the teacher network form a semi-supervised architecture for training: the student network is trained using labeled DAS data, the learnable weights are updated through back-propagation, and the teacher network is gradually updated by applying exponential moving average; the teacher network predicts the weakly enhanced unlabeled DAS data, and the prediction results that are higher than the set event classification confidence threshold are added to the training as DAS event classification pseudo labels, so as to train the DAS event classification branch on the unlabeled DAS data, and iteratively update the weights of the student network; The predicted results below the set interval regression variance threshold are used as pseudo interval labels of DAS events to train the interval DAS event regression branch on unlabeled DAS data, and the weights of the student network are iteratively updated; the student network is optimized through supervision loss, classification loss of DAS event classification pseudo labels and regression loss of DAS event pseudo interval labels, and the parameters of the teacher network are only updated by the student network through exponential moving average; The DAS data is passed through the backbone network for feature extraction in turn, and the obtained feature map is upsampled and fused through the neck network to obtain a preliminary prediction result. Finally, the final DAS event classification prediction result and DAS event positioning prediction result are obtained through the DAS event classification branch and DAS event regression branch of the detection head respectively.

2. The semi-supervised DAS event detection method according to claim 1, characterized in that: By building a distributed optical fiber sensing data acquisition platform, the monitoring signal of the DAS event is obtained to obtain the DAS data.

3. The semi-supervised DAS event detection method according to claim 1, characterized in that: The DAS data is sequentially passed through the backbone network for feature extraction, specifically including: The DAS data undergoes downsampling operation N1 to output feature map group C1, C1 undergoes multiple feature fusion downsampling operations N2 in turn to output feature map groups C2, C3, and C4, and C4 undergoes convolutional attention fusion operation N3 to output feature map group C5.

4. The semi-supervised DAS event detection method according to claim 3, characterized in that: The downsampling operation N1 consists of a one-dimensional convolutional layer Conv1d, batch normalization BatchNorm1d and activation function SiLU; the feature fusion downsampling operation N2 consists of a downsampling operation N1 and a segmentation operation Split, a concatenation operation Concat and an addition operation Add; the convolutional attention fusion operation N3 consists of a downsampling operation N1, a maximum pooling layer Maxpooling, an attention module Attention and a feedforward neural network FFN.

5. The semi-supervised DAS event detection method according to claim 3, characterized in that: The feature map is obtained by upsampling the neck network and fusion of features to obtain a preliminary prediction result, which specifically includes: After the upsampling operation N4, the feature map group C5 is merged with the feature map group C4 by channel, and the feature fusion N5 outputs the feature map group F4. After the upsampling operation N4, F4 is merged with the feature map group C3 by channel, and the preliminary prediction result P3 is output through N5. After the downsampling operation N1, P3 is merged with F4 by channel, and the preliminary prediction result P4 is output through N5. After the downsampling operation N1, P4 is merged with C5 by channel, and the preliminary prediction result P5 is output through feature fusion N5.

6. The semi-supervised DAS event detection method according to claim 5, characterized in that: The feature fusion operation N5 is composed of a one-dimensional convolution layer, batch normalization and SiLU activation function, as well as a segmentation operation, a splicing operation, and an addition operation.

7. The semi-supervised DAS event detection method according to claim 5, characterized in that: The final event classification prediction result and event location prediction result obtained by the detection head specifically include: The preliminary prediction results P3, P4, and P5 are respectively subjected to multi-dimensional convolution and compound loss optimization operation N6 to obtain the final event classification prediction results and event location prediction results; the multi-dimensional convolution and compound loss optimization operation N6 consists of downsampling operation N1, depthwise separable convolution DWConv, and IOU loss, distribution focus loss, and binary cross entropy loss BCE.

8. The semi-supervised DAS event detection method according to claim 1, characterized in that: The student network is optimized by supervision loss, classification loss of DAS event classification pseudo labels and regression loss of DAS event pseudo interval labels, and the parameters of the teacher network are only updated by the student network through exponential moving average, specifically including: The labeled DAS data is weakly enhanced and then input into the student network for supervised training, and the supervision loss L is calculated. s ; in, represents the i-th labeled DAS data, is the classification loss of the DAS event classification label, is the regression loss of the DAS event interval label; The unlabeled DAS data are processed by strong enhancement and weak enhancement respectively. The weakly enhanced DAS data are input into the teacher network for prediction. The prediction results higher than the set event classification confidence threshold are used as DAS event classification pseudo labels, and the classification loss of the unlabeled DAS data is calculated. in, The i-th unlabeled DAS data, express The corresponding DAS event pseudo labels generated by the teacher network for classification; For the prediction results below the set interval regression variance threshold, they are used as pseudo interval labels for DAS events, and the regression loss of unlabeled DAS data is calculated. in express The corresponding pseudo interval labels of DAS events generated by the teacher network for regression; Finally, the total unsupervised loss is obtained In each iteration of the training phase, the student network passes through L s and L u The total loss L composed of weighted fusion is optimized: λ u Parameters contributing to controlling the unsupervised loss; The parameters of the teacher network are updated by the student network through exponential moving average, with weights θ s Student Model M s and weight θ t Teacher Model M t ,θ s and θ t The exponential moving average update process between is as follows: where the superscript i denotes the i-th iteration during training and α denotes the smoothing coefficient of the exponential moving average.

9. A semi-supervised DAS event detection system, characterized in that: The DAS event detection model used includes a core event detection network, a student network and a teacher network composed of the core event detection network; the core event detection network includes a backbone network, a neck network and a detection head; the semi-supervised DAS event detection system specifically includes: In the training module, the student network and the teacher network form a semi-supervised architecture for training: the student network is trained using labeled DAS data, the learnable weights are updated through back-propagation, and the teacher network is gradually updated by applying exponential moving average; the teacher network predicts the weakly enhanced unlabeled DAS data, and the predicted results above the set event classification confidence threshold are added to the training as DAS event classification pseudo-labels, so as to train the DAS event classification branch on the unlabeled DAS data, and iteratively update the weights of the student network; the predicted results below the set interval regression variance threshold are used as DAS event pseudo-interval labels, so as to train the interval DAS event regression branch on the unlabeled DAS data, and iteratively update the weights of the student network; the student network is optimized through the supervision loss, the classification loss of the DAS event classification pseudo-labels, and the regression loss of the DAS event pseudo-interval labels, and the parameters of the teacher network are only updated by the student network through exponential moving average; In the prediction module, the DAS data is extracted through the backbone network in sequence, and the obtained feature map is upsampled and fused through the neck network to obtain a preliminary prediction result. Finally, the final DAS event classification prediction result and DAS event positioning prediction result are obtained through the DAS event classification branch and DAS event regression branch of the detection head respectively.

Images