Distributed optical fiber temperature event identification method based on MFDS-GFNet network

Abstract

The application relates to the technical field of distributed optical fiber sensing signal recognition, and discloses a distributed optical fiber temperature event recognition method based on an MFDS-GFNet network, which aims to solve the problems of insufficient feature extraction, inflexible model architecture and difficulty in distinguishing multiple temperature events in the prior art distributed optical fiber temperature event recognition method.The method comprises the following steps: collecting distributed optical fiber original temperature data, preprocessing the original temperature data, extracting multiple features, and constructing an MFDS-GFNet network model to realize distributed optical fiber temperature event recognition.The application proposes a gating fusion mechanism, which automatically adjusts the importance weight of different features through a learnable neural network, significantly improves the recognition accuracy, and is especially suitable for distinguishing event types that are difficult to distinguish.The application improves the recognition accuracy of events with similar temperature characteristics, such as water baths and heating belts, to 100%.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of distributed optical fiber sensing signal recognition, and particularly relates to a distributed optical fiber temperature event recognition method based on a multi-feature double-flow gated fusion network (MFDS-GFNet network). BACKGROUND

[0002] As a new temperature monitoring method, the distributed optical fiber temperature sensing technology is based on the BOTDR system of Brillouin scattering. The BOTDR system utilizes the Brillouin scattering generated by the propagation of optical pulses in the optical fiber, and realizes distributed parameter measurement by measuring the relationship between the scattering light frequency shift and temperature and strain. Compared with the temperature monitoring system based on Raman scattering, the BOTDR system has the advantages of long measurement distance, high spatial resolution, and simultaneous measurement of temperature and strain, and has a wide application prospect in the fields of power cable monitoring, oil and gas pipeline leakage detection, and large-scale structure health monitoring.

[0003] The BOTDR system can simultaneously measure temperature and strain parameters. Although in the field of stress research, BOTDR has been widely used in the detection of physical phenomena such as galloping and icing, the characteristics of its temperature data and its performance in actual application environment still need further research in the aspect of temperature event recognition. The BOTDR temperature data presents time-space two-dimensional distribution characteristics, and different types of temperature events show unique patterns on this distribution. However, current researches usually focus on the detection of single event type, or only use limited data features, and fail to fully exploit the rich spatio-temporal characteristics of distributed optical fiber temperature data. In particular, for the characteristics of BOTDR temperature data, designing an efficient recognition algorithm, and in the aspects of simultaneously recognizing multiple temperature event types and integrating multi-dimensional feature information, the existing researches still have obvious deficiencies. The complex and diverse characteristics of temperature events bring challenges to automatic recognition, and it is necessary to design targeted analysis methods.

[0004] In recent years, machine learning and deep learning techniques have made breakthroughs in complex pattern recognition tasks, providing new solutions for distributed fiber sensing data analysis. In the field of vibration fiber sensing, Sun et al. proposed a CNN-BiLSTM network based on Bayesian optimization for distributed fiber vibration classification. This method achieved an average recognition accuracy of 98.76% on nine different categories of sensing data; Jiang et al. proposed a high-precision classification method based on Vision Transformer for identifying and classifying vibration events in the φ-OTDR system, which achieved a classification accuracy of 99.3% and a false alarm rate of 0.2% to 1.8%; in 2019, Shi et al. first applied a simplified version of GoogLeNet to event classification in φ-OTDR, achieving an accuracy of 96.67% in identifying five different types of events, and the retraining time for a new sensing setup was only 7 minutes; Gao et al. proposed a 2D-CNN structure based on channel and spatial attention mechanisms (CS-CNN), which achieved 2 times faster convergence and 11% higher accuracy than traditional 2D-CNN.

[0005] For distributed temperature fiber sensing, Araujo et al. applied autoencoders to DTS systems, enabling them to detect smaller leaks and have a lower false alarm rate; DATT et al. demonstrated the effectiveness of Bi-GRU in multi-temperature event classification; K. Kalli, et al. applied machine learning to fiber distributed sensing, particularly a variant based on the Transformer architecture for temperature data anomaly detection. The study used a commercial BOTDR distributed sensing system to extract temperature information from power line distribution networks and achieved spatio-temporal anomaly point prediction through attention mechanisms for monitoring potential cable fault conditions. This study demonstrated the potential of applying Transformer architecture to fiber temperature event recognition, providing a new solution for power system safety monitoring.

[0006] However, existing methods have some key problems in practical applications: first, multi-class temperature events are difficult to distinguish, especially for events with similar temperature characteristics; second, feature extraction is insufficient, and the multi-dimensional features of temperature data are not fully exploited; third, the model architecture is not flexible enough to adapt to different types of data input and environmental conditions. Therefore, how to design a deep learning model that can simultaneously consider raw temperature data and multiple physical features, with strong feature fusion capability, is the key to improving the accuracy of distributed fiber temperature event recognition. SUMMARY

[0007] The present application aims at the problems of insufficient feature extraction, inflexible model architecture and difficulty in distinguishing multi-class temperature events in existing distributed optical fiber temperature event recognition methods, and proposes a distributed optical fiber temperature event recognition method based on an MFDS-GFNet network model.

[0008] The distributed optical fiber temperature event recognition method based on the MFDS-GFNet network has the implementation process as follows:

[0009] Step one, collecting distributed optical fiber original temperature data by using a BOTDR system, pre-processing the original temperature data, and obtaining pre-processed temperature data;

[0010] Step two, multi-feature extraction;

[0011] Extracting multiple features from the pre-processed temperature data, and selecting the best feature set through feature importance analysis;

[0012] Step three, constructing an MFDS-GFNet network model, wherein the MFDS-GFNet network model comprises a double-flow network module and a gated fusion module;

[0013] In the training process, the pre-processed temperature data of step one and the best feature set extracted in step two are subjected to feature flow extraction through the double-flow network module, and the two feature flows are integrated through the gated fusion module to realize adaptive feature fusion; and the probability distribution of the temperature event is output through the classifier;

[0014] The parameters of the MFDS-GFNet network model are optimized by minimizing the cross-entropy loss function and the back propagation algorithm, and the training of the MFDS-GFNet network model is completed;

[0015] Step four, in the testing process, input the test data into the trained MFDS-GFNet network model, wherein the output of the network model is a five-dimensional vector corresponding to the prediction probability of five types of temperature events; the index position corresponding to the maximum prediction probability is taken as the event prediction label to realize automatic recognition of the temperature event type.

[0016] The present application has the following advantages:

[0017] Firstly, in the method of the present application, a feature extraction framework for distributed optical fiber temperature event recognition is designed, and multiple space-time features such as time gradient features, space gradient features and local statistical features are extracted from the original temperature data to accurately capture the feature information of different types of temperature events.

[0018] Second, in the MFDS-GFNet network model described in this invention, a dual-stream network structure is proposed. The mainstream network processes the raw temperature data, while the tributary network processes the extracted physical features. The complementary information flow improves the accuracy of distributed optical fiber temperature event identification, especially the ability to distinguish similar types of events.

[0019] Third, in the MFDS-GFNet network model described in this invention, a gated fusion mechanism is proposed, which automatically adjusts the importance weights of different features through a learnable neural network, significantly improving the recognition accuracy, especially for event types that are difficult to distinguish.

[0020] IV. Experiments demonstrate that the MFDS-GFNet network model of this invention achieves an accuracy of 99.74% on the identification task of five typical distributed optical fiber temperature events (background, water bath, heating zone, alcohol lamp, and freezing), which is 1.07 percentage points higher than the Transformer model and 1.53 percentage points higher than GoogLeNet. In particular, for events with similar temperature characteristics, such as water baths and heating zones, the identification accuracy is improved to 100%. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the distributed optical fiber temperature event identification method based on the MFDS-GFNet network model described in this invention.

[0022] Figure 2 This is a comparison chart of the local standard deviation, temporal gradient characteristics, and spatial gradient characteristics of two temperature events, namely water bath and heating zone, in this invention.

[0023] Figure 3 This is a detailed schematic diagram of the MFDS-GFNet dual-stream network architecture in this invention;

[0024] Figure 4 This is a schematic diagram of the internal structure of the Residual Unit, the basic unit of the ResNet network in this invention;

[0025] Figure 5 Here is a complete structure diagram of the ResNet network used for feature extraction;

[0026] Figure 6 Detailed structural diagram of the gating fusion mechanism for feature adaptive integration;

[0027] Figure 7 The bar chart shows the comparison of the accuracy of four different fusion methods in this invention—simple splicing, additive fusion, attention fusion, and gating fusion—in recognizing various temperature events.

[0028] Figure 8The confusion matrix effect diagram of the temperature event identification of each identification model is shown in the figure, wherein (a) is the effect diagram of 1DCNN, (b) is the effect diagram of 2DCNN, (c) is the effect diagram of 2DCNN-BiLSTM, (d) is the effect diagram of GoogLeNet, (e) is the effect diagram of Transformer, and (f) is the effect diagram of the method MFDS-GFNet.

[0029] Figure 9 The accuracy rate and loss function change curve comparison diagram of the training process of the six different models in the application is shown in the figure, wherein (a) is the accuracy rate change curve effect diagram, and (b) is the loss function change curve effect diagram. DETAILED DESCRIPTION

[0030] Specific implementation Figures 1 to 9 The present embodiment is based on the distributed optical fiber temperature event identification method of the MFDS-GFNet network, which fully analyzes the time characteristics and spatial characteristics of the BOTDR temperature data, and improves the identification accuracy of the temperature event by mining the multi-dimensional feature relationship. As shown in the figure, Figure 1 The identification method of the present embodiment includes data preprocessing, multi-feature extraction, double-flow network structure and gate feature fusion steps. The specific implementation process is as follows:

[0031] Step 1: Use the BOTDR system to collect the original temperature data of the distributed optical fiber, and the preprocessing of the original temperature data includes ROI extraction, noise filtering and data standardization;

[0032] The original temperature data is stored in a two-dimensional format, with each row representing a time point sampling and each column representing the temperature value of different position points on the optical fiber. In order to better analyze the space-time characteristics of the temperature event, the data is converted into a matrix R with dimensions [T x L], where T = 100 represents the number of time sampling points, L = 2000 represents the number of spatial sampling points of the optical fiber, and the spatial resolution is 1 meter per point.

[0033] The specific process of preprocessing the original temperature data is as follows:

[0034] Step 1-1. ROI extraction: In the 2000-meter optical fiber monitoring data, the temperature event may occur in any region, but due to the limitation of computing resources and the need to improve processing efficiency, it is necessary to extract the ROI including the temperature event from the full-length data. For each measurement data, an adaptive ROI extraction algorithm is used: first, calculate the temperature standard deviation of each spatial point l in the entire time series to quantify the temperature fluctuation degree:

[0035]

[0036] In the formula, σ2 (l) represents the variance of temperature data at spatial point l over the entire time series, T represents the total duration of the entire time series, t is the time point, R t,l represents the temperature measurement value at time point t, spatial point l, μ l represents the average temperature of spatial point l over the entire time series, and the calculation formula is Identify the area where the temperature fluctuation is significant, detect all possible temperature event areas based on the temperature standard deviation distribution, when multiple possible event areas are detected, use distance-based clustering to merge adjacent areas, and select the most significant temperature event area as the ROI center:

[0037] wherein

[0038] In the formula, c represents the position of the finally selected ROI center, C * represents the "most significant" cluster, which is selected by the maximum average significance score, |C * | represents the number of spatial points included in the cluster C * , l∈C * represents each spatial point l belonging to C * , C i represents the i-th cluster, and S(l) is the significance score (based on temperature standard deviation) at position l. Based on the center point c, a ROI with a fixed width w is extracted, and in the experiment, the ROI width is fixed at 200 meters, which ensures sufficient capture of the complete features of various temperature events, while significantly reducing the data processing amount and improving the calculation efficiency.

[0039] Step 1-2. Noise filtering: Temperature sensing signals often contain random noise, which affects the stability of event features. Median filtering is used to reduce random noise while preserving edge information; median filtering is performed in both time and space dimensions. This two-dimensional filtering method is effective for different types of temperature events, especially for events with clear boundary characteristics such as heating bands, which can well preserve boundary information while suppressing measurement noise.

[0040] Step 1-3. Data standardization: Standardize the filtered data to the range [0, 1] to eliminate the influence of absolute temperature differences under different environmental conditions. Through the above preprocessing steps, the original temperature data of BOTDR is converted into a higher quality [100x200] dimensional matrix (preprocessed temperature data X), which is used as the input of the subsequent feature extraction and deep learning model.

[0041] Step 2: Multi-feature extraction and selection;

[0042] Five different features are extracted from the pre-processed temperature data X, and three of them are selected to accurately characterize the properties of temperature events. The five features are local statistical features, time gradient, space gradient, time second derivative, and temperature change amplitude.

[0043] Local statistical features (local std): Reflect the statistical properties of local temperature distribution, using the sliding window method to calculate the local standard deviation of temperature, to analyze the degree of local temperature fluctuation. The calculation formula is as follows:

[0044]

[0045] Where N represents the size of the sliding window, μ represents the average temperature of the N temperature data points in the sliding window. T(x i ,t i ) represents the temperature value of the i-th data point in the window.

[0046] Time gradient feature (time gradient): represents the rate of change of temperature at a certain spatial point on the sensing optical fiber with respect to time, and applies a first-order time difference operator to calculate the temperature value difference between adjacent time frames. The calculation formula is as follows:

[0047]

[0048] Where represents the gradient of temperature change with respect to time at spatial point x and time point t, T(x,t) represents the temperature value of spatial point x at time t, and Δt represents a very small time increment, usually the time interval between two consecutive temperature measurements (a time step), represents the partial derivative of temperature T with respect to time t, i.e. the instantaneous rate of change of temperature with respect to time.

[0049] Time second derivative (time accel): represents the acceleration of temperature change at a certain spatial point on the sensing optical fiber, used to detect temperature mutation, cooling or heating rate of rapid change behavior. The calculation formula is as follows:

[0050]

[0051] Where represents the second derivative of temperature change with respect to time at spatial point x and time point t, i.e. the acceleration of temperature change, represents the second-order partial derivative of temperature T with respect to time t.

[0052] Space gradient feature (space gradient): represents the rate of change of temperature spatial distribution on the optical fiber at a certain time, and applies a first-order spatial difference operator to calculate the temperature value difference between adjacent spatial positions. The calculation formula is as follows:

[0053]

[0054] where represents the gradient of temperature with respect to spatial position at spatial point x and time point t, represents the partial derivative of temperature T with respect to spatial coordinate x, and Δx represents a very small spatial increment, usually the distance between two adjacent measurement points on the optical fiber (a spatial step).

[0055] Temperature change magnitude: The spatial and temporal gradients are fused to comprehensively measure the intensity of temperature change at a certain spatial point. The calculation formula is as follows:

[0056]

[0057] M(x, t) represents the total magnitude or intensity of temperature change at spatial point x and time point t. It comprehensively considers the changes in time and space.

[0058] In this embodiment, to determine the optimal feature combination, the present application learns from the Fisher discriminant analysis method, and quantifies the discrimination ability of each feature by calculating the ratio of between-class variance to within-class variance (F statistic). Through systematic comparative experiments, it is found that the combination of local statistical features, time gradient features and spatial gradient features performs best, with an average test accuracy of 99.54%. The finally selected feature set is denoted as Φ(X) = {local statistical features, time gradient, spatial gradient}, which is used as the input of the tributary network.

[0059] As Figure 2 shown, Figure 2 Three key feature comparisons of water bath and heating belt events are shown. From the local standard deviation feature, the water bath event shows a uniform low value distribution, indicating a consistent temperature field distribution; while the heating belt event shows a higher local fluctuation, especially in the boundary area. In terms of time gradient features, the water bath event shows a low-intensity random distribution change, while the heating belt event shows a significant horizontal belt-shaped high gradient area at certain time points, reflecting the phased characteristics of the heating belt current regulation. In terms of spatial gradient features, the water bath event has significant gradients (bright stripes) only at the boundary, and the internal region gradient is almost zero; the heating belt event has high internal gradient throughout the region, showing an irregular bright spot distribution pattern, indicating the uneven distribution of temperature in the spatial dimension.

[0060] Step 3: As Figure 3 shown, the MFDS-GFNet network model is constructed; the MFDS-GFNet network architecture of the present application is designed with a double-flow structure, including a double-flow network module and a gated fusion module;

[0061] The double-flow network module is composed of a main flow network and a branch flow network; the double-flow network module of MFDS-GFNet processes the original temperature data and the extracted features respectively. The main flow network receives the pre-processed temperature data X ∈ R(T×L) as input, and directly extracts the spatio-temporal features from the original temperature data through a residual network structure; the branch flow network processes the feature set Φ(X) = {local statistical features, time gradient features, spatial gradient features} extracted from the original temperature data, focusing on the change characteristics of temperature rather than the absolute value.

[0062] In the information flow processing process, the input temperature data X first passes through the main flow network f main to generate the main flow feature vector V main ∈R 128 ; at the same time, the feature set Φ(X) is input into the branch flow network f supp to generate the branch flow feature vector V supp ∈R 128 .

[0063] The gating fusion module adaptively integrates the main flow feature vector V main and the branch flow feature vector V supp to generate the fusion feature vector V fused ∈R 128 , and then outputs the probability distribution of five temperature events through the SoftMax classifier.

[0064] As shown in Figure 4 and Figure 5 , the ResNet backbone network; the main flow and branch flow networks in MFDS-GFNet both adopt the residual network (ResNet) structure. The core of ResNet is to introduce residual connection to solve the gradient vanishing problem in deep network training, and its residual unit can be represented as:

[0065] h l = F(h l-1 , W l ) + h l-1

[0066] where h l represents the output of the lth layer, and F(·) represents the residual mapping function, which is specifically implemented as:

[0067] F(h l-1 , W l ) = W l,2 · σ(BN(W l,1 · BN(h l-1 )))

[0068] where W l,1 and W l,2are the parameters of the convolutional layer, and σ denotes the ReLU activation function, and BN denotes the batch normalization operation.

[0069] In this embodiment, the network main part is composed of three consecutive residual blocks, each containing two residual units, with a total of 3x2x2+2=14 layers (including the initial convolution and the final fully connected). In each residual unit, a serialized batch normalization (BN), a ReLU activation function and a 3x3 convolution operation are applied. In particular, a small 3x3 convolution kernel is chosen instead of the commonly used 7x7 convolution kernel to better preserve the local variation patterns in the temperature data, which is particularly important for identifying heating band events with sharp boundaries. To improve computational efficiency, the channel number of the first residual block is set to 64, while the last two residual blocks are 128. The transition between the first residual block and the second residual block uses a convolution with a step size of 2 for downsampling, and the feature map size changes from (H, W) to (H / 2, W / 2), and the receptive field increases to 1.78 times the original (from The calculation is as follows, where r l is the receptive field of the lth layer, k is the size of the convolution kernel, s j is the step size of the jth layer). This structure design enables the network to simultaneously focus on the local details and global distribution characteristics of temperature changes, effectively distinguishing different types of temperature event patterns.

[0070] Another feature of temperature data is that different events exhibit different characteristics in the time and spatial dimensions. For example, a water bath exhibits uniform temperature distribution (entropy H = -∑ i p i log p i is low) macroscopically but has slight fluctuations microscopically; while a heating band exhibits clear boundaries and uneven internal distribution (characteristic gradient is large). The hierarchical structure and residual connection of ResNet enable the network to retain and integrate multi-scale features, combine shallow detail features with deep semantic features through skip connections, and provide more comprehensive temperature event representation.

[0071] To further improve the model's perception of key features, a channel attention mechanism (SE Block) is integrated into each residual unit.

[0072] Finally, the feature map is converted into a fixed-length feature vector through global average pooling:

[0073]

[0074] where f i,j ∈R 128 is the feature vector at position (i,j), and finally v∈R128 .

[0075] It is worth noting that the main stream and branch network share the same architecture design, but maintain independent parameter spaces. This design takes into account the essential differences in data distribution between temperature raw data and physical characteristics, allowing the two streams to learn optimal representations for these two types of data respectively.

[0076] As shown in Figure 6 , Figure 6 The gating fusion module in MFDS-GFNet is shown. This module receives the feature vector V main ∈R 128 of the main stream network and the feature vector V supp ∈R 128 of the branch network, and determines their weight distribution in the final feature representation through adaptive learning.

[0077] The calculation process of the gating fusion module can be represented as follows: first, the main stream feature vector V main and the branch feature vector V supp are concatenated in dimension to form a joint feature representation:

[0078] V concat =[V main ;V supp ]∈R 256

[0079] Then, the gating coefficient is generated through a fully connected layer and a Sigmoid activation function:

[0080] g=σ(W g ·V concat +b g )∈R 128

[0081] Where W g ∈R 128×256 and b g ∈R 128 are the learnable parameter matrix and bias vector, and σ is the Sigmoid activation function.

[0082] Finally, the gating coefficient is applied to element-level weighted fusion of the two feature vectors:

[0083] V fused =g⊙V main +(1-g)⊙V supp

[0084] Where ⊙ represents the Hadamard product, i.e., element-level multiplication operation.

[0085] Step 4, The training data uses a self-built BOTDR distributed optical fiber temperature event dataset, which contains five typical event types: background, water bath, heating band, alcohol lamp, and freezing. Each type has about 2000 samples, and the total number of samples is about 10000. To ensure the stability and generalization ability of model training, the dataset is divided into training set, validation set and test set in the ratio of 6:2:2.

[0086] In the training stage, the preprocessed normalized temperature data is used as the main stream network input, and the extracted three features (local standard deviation, time gradient and spatial gradient) are used as the branch network input. MFDS-GFNet uses a dual-stream structure to extract original and physical feature vectors, and performs feature weighting and integration through a gating fusion module. Finally, the Softmax classifier outputs the probability distribution of temperature events. The network optimizes all parameters by minimizing the cross-entropy loss function and combining the back propagation algorithm.

[0087] In the test stage, the test set samples are input into the trained MFDS-GFNet network, and the model output is a five-dimensional vector corresponding to the prediction probability of the five temperature events. The system uses the index position corresponding to the maximum probability as the event prediction label to realize automatic recognition of temperature event types.

[0088] This method finally realizes an end-to-end temperature event recognition process with preprocessed temperature data and multi-dimensional feature combination as input and event category label as output, significantly improving the ability to distinguish similar events (such as water bath and heating band), and having good real-time performance and robustness.

[0089] DETAILED DESCRIPTION Figure 2 、 Figures 7 to 9 This embodiment is an example of experimental verification of the distributed optical fiber temperature event recognition method based on MFDS-GFNet network described in embodiment one:

[0090] First, in this embodiment, all experiments are performed on a workstation with NVIDIA GeForce RTX 3090 GPU and 32GB memory, using PyTorch 1.9.0 deep learning framework. The training parameters of MFDS-GFNet and all comparison models (including 1DCNN, 2DCNN, 2DCNN-BiLSTM, GoogLeNet and Transformer) are set as follows: the batch size is set to 32, the initial learning rate is 0.001, the Adam optimizer (β1=0.9, β2=0.999) is used, the weight decay is 0.0001, and the training rounds are 50.

[0091] To optimize the training process, a cosine annealing learning rate scheduling strategy was adopted, which made the learning rate larger in the early stage of training to accelerate convergence, and gradually reduced in the later stage of training to improve accuracy. To prevent overfitting, the following techniques were applied: early stopping strategy (patience = 20), which stopped training when the validation set loss did not improve for 20 consecutive epochs; data augmentation techniques, including, in the order of application: random rotation (±10°), random scaling (0.9-1.1 times), random horizontal flip (probability 0.5), and random noise addition (Gaussian noise, σ = 0.01).

[0092] In this embodiment, the precision, recall, and F1 score are used as three standard indicators to evaluate the performance of the results of the intrusion event identification method, which are calculated according to TP (true positive), FP (false positive), TN (true negative), and FN (false negative) in the confusion matrix, as follows

[0093]

[0094] Secondly, the rationality of MFDS-GFNet is verified; compared with the performance of 1DCNN, 2DCNN, 2DCNN-BiLSTM, GoogLeNet, and Transformer commonly used in current temperature event identification, it is verified that MFDS-GFNet has better identification performance.

[0095] As Figure 2 shown, in order to more intuitively observe the differences of different temperature events in key features, this figure shows the comparison of water bath and heating belt events in three core features (local standard deviation, time gradient, and space gradient). From the local standard deviation feature, the water bath event shows a uniform low value distribution, indicating a consistent temperature field distribution; while the heating belt event shows higher local fluctuations, especially in the boundary area. In the time gradient feature, the water bath event shows a low-intensity random distribution of changes, while the heating belt event shows obvious horizontal band-shaped high gradient areas at certain time points, reflecting the phased characteristics of the heating belt current regulation. In the space gradient feature, the water bath event has significant gradient (bright stripes) only at the boundary, and the internal area gradient is almost zero; the heating belt event has high gradient in the entire internal area, showing an irregular bright spot distribution pattern. The differences in these features provide a reliable basis for the automatic identification of temperature events in distributed optical fiber sensing systems.

[0096] To verify the effectiveness of the feature selection method proposed in this embodiment, a system evaluation is performed on different feature combinations. The results show that the selected three-feature combination (local standard deviation + time gradient + space gradient) has a recognition accuracy of 99.74%, which is much higher than that of a single feature and other feature combinations. In particular, compared with using all five features (98.85%), the selected three-feature combination performs better, indicating that the complementarity of features is more important than the number of features.

[0097] As shown in Figure 7 To verify the superiority of the proposed gating fusion mechanism, it is compared with three traditional fusion strategies: simple concatenation (Concat), additive fusion (Add), and attention fusion (Attention). The gating fusion (Gate fusion) shows the best performance, with an overall accuracy of 99.74%, which is 0.56 percentage points higher than the second best method of concatenation (99.18%). In particular, for heating belt events, the accuracy of the gating fusion is 100%, while other methods perform poorly on this type of event (concatenation 97.92%, addition 96.86%, attention 97.38%). This indicates that the gating fusion can more effectively handle the feature relationships of complex events, especially for heating belt events with high spatiotemporal variation characteristics.

[0098] To evaluate the contribution of each component in the MFDS-GFNet network model, a series of ablation experiments are performed. The results show that the ResNet base model (without features, without SE module, without gating fusion) has an accuracy of 95.46%, and adding a single feature can improve the accuracy to 97.79%-98.80%, with the space gradient feature being the best (98.80%). When the three key features are combined, the accuracy reaches 99.18%. Although the SE module alone has no significant effect (96.88%), it plays a synergistic role in the complete model. The complete MFDS-GFNet model (including all components) achieves the best performance of 99.74%, proving the rationality of the design of each component.

[0099] To comprehensively evaluate the effectiveness of MFDS-GFNet, it is compared with current mainstream deep learning models, including 1DCNN, 2DCNN, 2DCNN-BiLSTM, GoogLeNet, and Transformer. Table 1 shows the performance comparison results of different models.

[0100] Table 1

[0101]

[0102] Among all evaluation indicators, the MFDS-GFNet network model in the embodiment shows obvious advantages, and the test accuracy rate reaches 99.74%, which is 1.07 percentage points higher than the second best Transformer model (98.67%) and 1.53 percentage points higher than the widely used GoogLeNet (98.21%). In terms of precision, recall and F1 score and other indicators, the MFDS-GFNet also maintains a high level and balance, indicating that the model not only has high overall accuracy, but also performs stably in the identification of various events, without a significant decrease in the performance of some event types.

[0103] As shown in Figure 8 , the recognition ability of different models for five temperature events can be more clearly observed through the confusion matrix. Among the compared deep learning models, 1DCNN, as the most basic time series model, has obvious confusion in multiple event types, especially between water bath and heating band events, with a mutual misclassification rate as high as 8-10%, which indicates that considering only the time dimension features cannot effectively distinguish these two types of events with similar temperature ranges but different spatial distributions. With the increase of model complexity, the recognition performance gradually improves: 2DCNN reduces the misclassification rate between water bath and heating band events to 4-5%; 2DCNN-BiLSTM further reduces the confusion between water bath and heating band (about 3%); GoogLeNet reduces the confusion rate to about 2%; Transformer reduces it to about 1.5%. In contrast, MFDS-GFNet shows excellent discrimination ability, and the confusion matrix shows that only a small amount of water bath samples (about 0.5%) are misclassified as heating band, and the recognition of other event types reaches or approaches 100% accuracy.

[0104] As shown in Figure 9 , it can be seen that among the six deep learning methods, MFDS-GFNet has the fastest convergence speed, and its validation accuracy quickly rises above 95% in the early training stage (about 5 epochs), while the accuracy of other models in the same training stage is generally below 85%. In particular, MFDS-GFNet only needs about 15 epochs to reach an accuracy of more than 98%, while Transformer needs about 25 epochs, and GoogLeNet and 2DCNN-BiLSTM need 30+ epochs. From the loss curve, MFDS-GFNet not only has the fastest decline speed, but also has the smoothest curve, and the fluctuation in the later training stage is significantly smaller than that of other models. This indicates that the training process of MFDS-GFNet is more stable and less susceptible to random fluctuations.

[0105] The fast convergence and training stability of the MFDS-GFNet benefit from its multi-feature fusion architecture, which enables the model to utilize multiple complementary features simultaneously and accelerate the search for the optimal solution. In particular, the adaptive nature of the gating fusion mechanism allows the model to dynamically adjust the importance of different features during training, quickly focusing on the most discriminative feature combinations. This property is highly valuable in practical applications, significantly reducing model training time and resource consumption.

[0106] In terms of computational efficiency, although the MFDS-GFNet introduces multi-feature extraction and gating fusion mechanisms, its parameter quantity (7.88M) is lower than that of the Transformer (9.46M) and 2DCNN-BiLSTM (8.32M), and slightly higher than that of the 2DCNN (6.32M), indicating that the model has achieved a good balance between complexity and performance. In terms of inference time, the MFDS-GFNet processing time per sample is 0.40 milliseconds, which is higher than that of simple models such as 1DCNN (0.24 milliseconds) and 2DCNN (0.30 milliseconds), but lower than that of the Transformer (0.44 milliseconds), fully meeting the real-time processing requirements in practical applications.

[0107] Finally, through the above experimental verification, the MFDS-GFNet network model of the present application fully utilizes the spatio-temporal information of distributed optical fiber temperature data, and can better and faster identify various temperature events, especially for events with similar temperature characteristics (such as water bath and heating belt), with excellent discrimination ability. The experimental results show that, on the five typical temperature event (background, water bath, heating belt, alcohol lamp and freezing) recognition task, the MFDS-GFNet achieves a high accuracy of 99.74%, and achieves almost perfect recognition (accuracy up to 100%) on the difficult-to-distinguish water bath and heating belt events. At the same time, the model shows faster convergence speed and more stable training process, as well as moderate computing resource requirement, with high practical value.

[0108] The technical features of the above-described embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described, but it should be considered that any combination of the technical features is within the scope of the present disclosure as long as the combination does not result in contradictions.

[0109] The above-described embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A distributed optical fiber temperature event identification method based on MFDS-GFNet network, characterized by: The implementation process of this method is as follows: Step 1: Use the BOTDR system to collect raw temperature data from distributed optical fibers, preprocess the raw temperature data, and obtain preprocessed temperature data. Step 2: Multi-feature extraction; Multiple features are extracted from the preprocessed temperature data, and their importance is assessed. Analyze and select the optimal feature set; Step 3: Construct the MFDS-GFNet network model, which consists of a two-stream network module and a gated fusion module; During training, the temperature data preprocessed in step one and the optimal feature set extracted in step two are used for feature stream extraction through a dual-stream network module, and the two feature streams are integrated through a gated fusion module to achieve adaptive feature fusion; and the probability distribution of temperature events is output through a classifier. The parameters of the MFDS-GFNet network model are optimized by minimizing the cross-entropy loss function and using the backpropagation algorithm, thus completing the training of the MFDS-GFNet network model. The dual-stream network module consists of a mainstream network and a tributary network. The mainstream network receives the temperature data after preprocessing in step one, and extracts spatiotemporal features through the residual network structure to obtain the mainstream feature vector; The tributary network receives the optimal feature set obtained in step two and generates tributary feature vectors. The gated fusion module adaptively integrates the mainstream feature vector and the tributary feature vector to generate a fused feature vector, and then outputs the probability distribution of temperature events through the SoftMax classifier. Both the mainstream and tributary networks adopt the ResNet network structure. The ResNet network structure consists of three residual blocks, each of which includes two residual units. Each residual unit integrates a channel attention mechanism and converts the feature map into a fixed-length feature vector through global average pooling. The gated fusion module concatenates the mainstream feature vector and the tributary feature vector in terms of dimension to form a joint feature representation. It then generates a gate coefficient g through a fully connected layer and a Sigmoid activation function. Finally, it uses the gate coefficient to perform element-wise weighted fusion of the two feature vectors to obtain a fused feature vector. Step 4: During the testing process, the test data is input into the trained MFDS-GFNet network model. The network model outputs a five-dimensional vector, which corresponds to the predicted probability of five types of temperature events. The index position corresponding to the maximum predicted probability is used as the event prediction label to realize the automatic identification of temperature event types.

2. The distributed optical fiber temperature event identification method based on MFDS-GFNet network according to claim 1, characterized in that: In step one, the preprocessing of the raw temperature data is as follows: Step 1: Extract the ROI from the raw temperature data, calculate the temperature standard deviation of each spatial point in the time series, detect temperature event regions based on the distribution of temperature standard deviation, and select the temperature event regions as the ROI centers. Steps 1 and 2: Perform median filtering in both time and space dimensions to filter noise from the original temperature data; Step 13: Standardize the noise-filtered data to the range [0,1] to obtain the preprocessed temperature data X.

3. The distributed optical fiber temperature event identification method based on MFDS-GFNet network according to claim 2, characterized in that: In step one, when multiple event regions are detected, distance-based clustering is used to merge adjacent regions, and the most significant temperature event region is selected as the ROI center, expressed by the following formula: in In the formula, C i For the i-th cluster, C * For the most significant cluster, S(l) is the significance score at position l. Based on the center point c, a ROI with a fixed width w is extracted.

4. The distributed optical fiber temperature event identification method based on MFDS-GFNet network according to claim 1, characterized in that: In step two, the extracted features include local statistical features, temporal gradient features, spatial gradient features, second time derivative, and temperature change amplitude. The discriminative power of each feature is quantified by calculating the ratio of between-class variance to within-class variance, and the optimal feature set is selected.

5. The distributed optical fiber temperature event identification method based on MFDS-GFNet network according to claim 1, characterized in that: The joint feature representation V concat It can be expressed as follows: V concat =[V main ;V supp ]∈R 256 In the formula, V main V is the mainstream feature vector. supp The characteristic vector of the tributary; The gating coefficient g is expressed by the following formula: g=σ(W g ·V concat +b g )∈R 128 In the formula, W g and b g Let be the learnable parameter matrix and bias vector, and σ be the Sigmoid activation function; Fusion feature vector V fused It can be expressed as follows: V fused =g⊙V main +(1-g)⊙V supp In the formula, ⊙ represents the Hadamard product, which is an element-wise multiplication operation.

Images