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

Abstract

The invention discloses a distributed optical fiber temperature event identification method based on an MFDS-GFNet network, relates to the technical field of distributed optical fiber sensing signal identification, and aims to solve the problems that an existing distributed optical fiber temperature event identification method is insufficient in feature extraction, not flexible enough in model architecture, difficult to distinguish multiple types of temperature events and the like. The method comprises the following steps: acquiring original temperature data of a distributed optical fiber, and preprocessing the original temperature data; extracting multiple features; and constructing an MFDS-GFNet network model to realize distributed optical fiber temperature event identification. According to the method, a gating fusion mechanism is provided, importance weights of different features are automatically adjusted through a learnable neural network, and the recognition accuracy is remarkably improved, especially for event types which are difficult to distinguish. According to the invention, the recognition accuracy of events with similar temperature characteristics, such as water bath and heating tape, is improved to 100%.

Description

Technical Field

[0001] The present invention relates to the technical field of distributed optical fiber sensing signal recognition, and in particular to a distributed optical fiber temperature event recognition method based on a multi-feature dual-stream gated fusion network (MFDS-GFNet network). Background Art

[0002] Distributed fiber optic temperature sensing technology is an emerging temperature monitoring method. Brillouin scattering-based BOTDR systems exploit Brillouin scattering generated when light pulses propagate through optical fibers. By measuring the relationship between the scattered light frequency shift and temperature and strain, they achieve distributed parameter measurement. Compared to Raman scattering-based temperature monitoring systems, these systems offer advantages such as longer measurement distances, higher spatial resolution, and the ability to simultaneously measure temperature and strain. They hold broad application prospects in areas such as power cable monitoring, oil and gas pipeline leak detection, and large-scale structural health monitoring.

[0003] The BOTDR system is capable of simultaneously measuring temperature and strain parameters. Although BOTDR has been widely used in the field of stress research to detect physical phenomena such as dancing and icing, its temperature data characteristics and its performance in actual application environments still require in-depth research in temperature event recognition. BOTDR temperature data exhibits a two-dimensional time-space distribution characteristic, and different types of temperature events show unique patterns in this distribution. However, current research generally focuses on the detection of a single event type, or only utilizes limited data features, failing to fully explore the rich spatiotemporal characteristics of distributed fiber optic temperature data. In particular, existing research still has significant deficiencies in designing efficient recognition algorithms based on the characteristics of BOTDR temperature data, as well as in simultaneously identifying multiple temperature event types and integrating multi-dimensional feature information. The complex and diverse characteristics of temperature events pose challenges to automatic recognition, requiring the design of targeted analysis methods.

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

[0005] For distributed temperature fiber optic sensing, Araujo et al. applied autoencoders to DTS systems, enabling them to detect smaller vulnerabilities with lower false negative rates. DATT et al. demonstrated the effectiveness of Bi-GRU in classifying multiple temperature events. K. Kalli et al. applied machine learning to fiber optic distributed sensing, specifically using a variant of the Transformer architecture for temperature data anomaly detection. This study used a commercial BOTDR distributed sensing system to extract temperature information from the power line distribution network. Using an attention mechanism, they predicted spatiotemporal anomalies for monitoring potential cable faults. This research demonstrates the potential of applying the Transformer architecture to fiber optic temperature event identification, providing a new solution for power system security monitoring.

[0006] However, existing methods face several key challenges in practical applications: first, it is difficult to distinguish between multiple types of temperature events, especially those with similar temperature characteristics; second, feature extraction is insufficient, failing to fully exploit the multidimensional characteristics of temperature data; and third, the model architecture is inflexible, making it difficult to adapt to different types of data input and environmental conditions. Therefore, designing a deep learning model that can simultaneously consider raw temperature data and multiple physical characteristics, with powerful feature fusion capabilities, is key to improving the accuracy of distributed fiber optic temperature event recognition. Summary of the Invention

[0007] In response to the problems existing in existing distributed optical fiber temperature event recognition methods, such as insufficient feature extraction, inflexible model architecture, and difficulty in distinguishing multiple types of temperature events, the present invention proposes a distributed optical fiber temperature event recognition method based on the MFDS-GFNet network model.

[0008] A distributed optical fiber temperature event recognition method based on the MFDS-GFNet network is implemented as follows:

[0009] Step 1: Use the BOTDR system to collect the original temperature data of the distributed optical fiber, pre-process the original temperature data, and obtain the pre-processed temperature data;

[0010] Step 2: Multi-feature extraction;

[0011] Extract multiple features from the preprocessed temperature data and select the best feature set through feature importance analysis;

[0012] Step 3: Construct the MFDS-GFNet network model, which consists of a dual-stream network module and a gated fusion module;

[0013] During the training process, the temperature data preprocessed in step 1 and the optimal feature set extracted in step 2 are extracted through a dual-stream network module for feature stream extraction. The two feature streams are then integrated through a gated fusion module to achieve adaptive feature fusion. The probability distribution of temperature events is then output through a classifier.

[0014] The training of the MFDS-GFNet network model is completed by optimizing the parameters of the MFDS-GFNet network model by minimizing the cross-entropy loss function and using the back-propagation algorithm.

[0015] Step 4: During the test 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 automatic recognition of the temperature event type.

[0016] Beneficial effects of the present invention:

[0017] 1. In the method of the present invention, a feature extraction framework for distributed optical fiber temperature event identification is designed to extract multiple spatiotemporal features such as time gradient features, spatial gradient features, and local statistical features from the original temperature data, accurately capturing the characteristic information of different types of temperature events.

[0018] 2. In the MFDS-GFNet network model described in the present invention, a dual-stream network structure is proposed. The mainstream network processes the original temperature data, and the tributary network processes the extracted physical features. The complementary information flows improve the accuracy of distributed optical fiber temperature event recognition, especially the ability to distinguish similar types of events.

[0019] 3. In the MFDS-GFNet network model described in the present invention, a gated fusion mechanism is proposed, which automatically adjusts the importance weights of different features through a learnable neural network, significantly improving recognition accuracy, especially for event types that are difficult to distinguish.

[0020] Fourth, experiments demonstrated that the MFDS-GFNet network model achieved an accuracy of 99.74% in the recognition task of five typical distributed fiber optic temperature events (background, water bath, heating zone, alcohol lamp, and freezing), 1.07 percentage points higher than the Transformer model and 1.53 percentage points higher than Google LeNet. In particular, for events with similar temperature characteristics, such as water baths and heating zones, the recognition accuracy increased to 100%. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of the distributed optical fiber temperature event identification method based on the MFDS-GFNet network model of the present invention;

[0022] Figure 2 This is a comparison diagram of the local standard deviation, time gradient characteristics, and spatial gradient characteristics of the two temperature events of the water bath and the heating zone in the present invention;

[0023] Figure 3 Schematic diagram of the detailed architecture of the MFDS-GFNet dual-stream network in the present invention;

[0024] Figure 4 Schematic diagram of the internal structure of the basic unit Residual Unit of the ResNet network in the present invention;

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

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

[0027] Figure 7 This is a bar chart comparing the recognition accuracy of various temperature events using the four different fusion methods of simple splicing, additive fusion, attention fusion, and gated fusion in this invention;

[0028] Figure 8The confusion matrix effect diagrams for temperature event recognition are compared among the recognition models; among them, (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 MFDS-GFNet of this method.

[0029] Figure 9 Comparison diagram of the accuracy and loss function change curves of the training process of six different models in the present invention; among them, (a) is the accuracy change curve effect diagram, and (b) is the loss function change curve effect diagram. DETAILED DESCRIPTION

[0030] Specific implementation method 1. Combination Figures 1 to 9 This embodiment describes a distributed optical fiber temperature event recognition method based on the MFDS-GFNet network. This method fully analyzes the temporal and spatial characteristics of BOTDR temperature data and mines its multi-dimensional feature relationships to improve the recognition accuracy of temperature events. Figure 1 As shown, the recognition method of this embodiment includes steps such as data preprocessing, multi-feature extraction, dual-stream network structure and gated feature fusion. Its specific implementation process is:

[0031] Step 1: Using a BOTDR system to collect raw temperature data of the distributed optical fiber, the raw temperature data is preprocessed including region of interest (ROI) extraction, noise filtering and data normalization;

[0032] The raw temperature data is stored in a two-dimensional format, with each row representing a sample at a specific time point and each column representing the temperature value at a different location on the fiber. To better analyze the spatiotemporal characteristics of temperature events, the data is converted into a matrix R with dimensions [T × L], where T = 100 represents the number of time sampling points and L = 2000 represents the number of spatial sampling points on the fiber. 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: Temperature events can occur anywhere in the 2000-meter fiber monitoring data. However, due to computational resource limitations and the need to improve processing efficiency, it is necessary to extract the region of interest (ROI) containing the temperature event from the full-length data. For each measurement, an adaptive ROI extraction algorithm is used: First, the temperature standard deviation of each spatial point l over the entire time series is calculated to quantify the degree of temperature fluctuation:

[0035]

[0036] Where, σ2 (l) represents the variance of the temperature data at the spatial point l in the entire time series, T represents the total length of the entire time series, t is the time point, R t,l represents the temperature measurement value at time point t and space point l, μ l Represents the average temperature of the spatial point l over the entire time series, and its calculation formula is Identify areas with significant temperature fluctuations. Based on the temperature standard deviation distribution, detect all possible temperature event areas. 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] in

[0038] Where c represents the position of the final selected ROI center, C * represents the “most significant” cluster, which is selected by the maximum average significance score, |C * | represents cluster C * The number of spatial points contained in (i.e., the size of the cluster), l∈C * Indicates that it belongs to C * For every spatial point l, C i Denotes the i-th cluster, and S(l) is the significance score at location l (based on the temperature standard deviation). Based on the center point c, a fixed width w ROI is extracted. In the experiment, the ROI width is fixed at 200 meters to ensure that it can capture the complete characteristics of various temperature events. At the same time, it significantly reduces the amount of data processing and improves computational 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. This dual-dimensional filtering method is effective for different types of temperature events, especially those with distinct boundary features, such as heating zones. It can effectively preserve boundary information while suppressing measurement noise.

[0040] Step 1-3. Data normalization: The filtered data is normalized to the range of [0, 1] to eliminate the influence of absolute temperature differences under different environmental conditions. Through the above preprocessing steps, the raw temperature data of the BOTDR is converted into a higher quality [100 × 200] dimensional matrix (preprocessed temperature data X) as the input for subsequent feature extraction and deep learning models.

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

[0042] Five different features are extracted from the preprocessed temperature data X, and three of them are selected to accurately characterize the characteristics of the temperature event. The five features are local statistical characteristics, time gradient, spatial gradient, time second derivative and temperature change amplitude.

[0043] Local statistical characteristics (local std): reflects the statistical characteristics of the local temperature distribution. The sliding window method is used to calculate the local standard deviation of the 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, and μ represents the average temperature of the N temperature data points in the sliding window. i ,t i ) represents the temperature value of the i-th data point in the window.

[0046] Time gradient characteristic (time gradient): Indicates the rate of change of temperature of a spatial point on the sensing fiber over time. It is calculated by applying a first-order time difference operator and the temperature difference between adjacent time frames. The calculation formula is as follows:

[0047]

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

[0049] Time second derivative (time accel): Indicates the acceleration of the temperature change at a spatial point on the sensing fiber. It is used to detect sudden temperature changes, rapid changes in cooling or heating rates. The calculation formula is as follows:

[0050]

[0051] in Represents the second-order derivative of the temperature at the spatial point x and time point t, that is, the acceleration of the temperature change. represents the second-order partial derivative of temperature T with respect to time t.

[0052] Spatial gradient: This indicates the rate of change of the spatial distribution of temperature on the optical fiber at a certain moment. It is calculated by applying a first-order spatial difference operator and the temperature difference between adjacent spatial locations. The calculation formula is as follows:

[0053]

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

[0055] Temperature change magnitude (magnitude): It integrates spatial and temporal gradients 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 the temperature change at a spatial point x and a time point t. It combines changes in time and space.

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

[0059] like Figure 2 As shown, Figure 2 The comparison of three key characteristics of water bath and heating belt events is shown. From the perspective of local standard deviation characteristics, water bath events show a uniform low-value distribution, indicating a consistent temperature field distribution; while heating belt events show higher local fluctuations, especially in the boundary area. In terms of time gradient characteristics, water bath events show changes in low-intensity random distribution, while heating belt events show obvious horizontal banded high-gradient areas at certain time points, reflecting the stage-specific nature of heating belt current regulation. In terms of spatial gradient characteristics, water bath events have significant gradients (bright stripes) only at the boundaries, and the gradient in the internal area is almost zero; heating belt events have high gradients throughout the entire area, showing an irregular bright spot distribution pattern, indicating an uneven distribution of temperature in the spatial dimension.

[0060] Step 3: If Figure 3 As shown, the MFDS-GFNet network model is constructed; the MFDS-GFNet network architecture of the present invention designs a dual-stream structure, including a dual-stream network module and a gated fusion module;

[0061] The dual-stream network module consists of a main network and a tributary network. The dual-stream network module of MFDS-GFNet processes raw temperature data and extracted features separately. The main network receives preprocessed temperature data X∈R(T×L) as input and directly extracts spatiotemporal features from the raw temperature data using a residual network structure. The tributary network processes the feature set Φ(X) = {local statistical features, temporal gradient features, spatial gradient features} extracted from the raw temperature data, focusing on the changing characteristics of temperature rather than absolute values.

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

[0063] The gated fusion module adaptively integrates the mainstream feature vector V main and tributary characteristic vector V supp , generate fusion feature vector V fused ∈R 128 , and then the probability distribution of five temperature events is output through the SoftMax classifier.

[0064] like Figure 4 and Figure 5 As shown in the figure, the ResNet backbone network; the mainstream and tributary networks in MFDS-GFNet all adopt the residual network (ResNet) structure. The core of ResNet is to introduce residual connections to solve the gradient vanishing problem in deep network training. Its residual unit can be expressed 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 follows:

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

[0068] Here W l,1 and W l,2is the convolutional layer parameter, σ represents the ReLU activation function, and BN represents the batch normalization operation.

[0069] In this embodiment, the main part of the network consists of three consecutive residual blocks, each residual block contains two residual units, and the total number of layers is 3×2×2+2=14 layers (including the initial convolution and the final full connection). In each residual unit, serialized batch normalization (BN), ReLU activation function and 3×3 convolution operation are applied. In particular, a small 3×3 convolution kernel is selected instead of the commonly used 7×7 convolution kernel to better preserve the local change pattern in the temperature data, which is especially important for identifying hot zone events with sharp boundaries. To improve computational efficiency, the number of channels of the first residual block is set to 64, and the number of channels of the next two residual blocks is 128. The transition between the first residual block and the second residual block uses a convolution with a step size of 2 for downsampling. The feature map size changes from (H, W) to (H / 2, W / 2), and the receptive field increases to 1.78 times the original (by Calculated, where r l is the receptive field of the lth layer, k is the convolution kernel size, s j This structural design enables the network to focus on both the local details and global distribution characteristics of temperature changes, effectively distinguishing different types of temperature event patterns.

[0070] Another characteristic of temperature data is that different events show different characteristics in time and space. For example, a water bath shows a uniform temperature distribution on a macro scale (entropy value H = -∑ i p i log p i The heating zone has a clear boundary and an uneven internal distribution (characteristic gradient The hierarchical structure and residual connections of ResNet enable the network to retain and integrate multi-scale features. It combines shallow detail features with deep semantic features through skip connections, providing a more comprehensive representation of temperature events.

[0071] In order to further improve the model's ability to perceive 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 eigenvector at position (i, j), and finally v∈R128 .

[0075] It is noteworthy that while the mainstream and tributary networks share the same architectural design, they maintain independent parameter spaces. This design takes into account the essential differences in the data distribution of raw temperature data and physical features, enabling the two streams to learn optimal representations for these two types of data.

[0076] like Figure 6 As shown, Figure 6 The gated fusion module in MFDS-GFNet is shown. This module receives the feature vector V of the mainstream network. main ∈R 128 and the characteristic vector V of the tributary network supp ∈R 128 , and their weight distribution in the final feature representation is determined through adaptive learning.

[0077] The calculation process of the gated fusion module can be expressed as follows: First, the mainstream feature vector V main and tributary characteristic vector V supp Concatenate in dimension to form joint feature representation:

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

[0079] Then the gating coefficients are generated through the fully connected layer and the Sigmoid activation function:

[0080] g=σ(W g ·V concat +b g )∈R 128

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

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

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

[0084] Where ⊙ represents the Hadamard product, which is an element-wise multiplication operation.

[0085] In step 4, the training data used a self-built BOTDR distributed fiber temperature event dataset. This dataset contains five typical event types: background, water bath, heating zone, alcohol lamp, and freezing. Each type contains approximately 2,000 samples, for a total of approximately 10,000 samples. To ensure the stability and generalization of model training, the dataset was divided into training, validation, and test sets in a 6:2:2 ratio.

[0086] During the training phase, preprocessed, standardized temperature data serves as the main network input, while the three extracted features (local standard deviation, temporal gradient, and spatial gradient) serve as the tributary network input. MFDS-GFNet utilizes a dual-stream architecture to extract raw and physical feature vectors, respectively. A gated fusion module performs weighted feature integration, ultimately outputting the probability distribution of temperature events through a softmax classifier. The network optimizes all parameters by minimizing the cross-entropy loss function and incorporating a backpropagation algorithm.

[0087] During the testing phase, the test set samples were fed into the trained MFDS-GFNet network. The model output was a five-dimensional vector, corresponding to the predicted probabilities of the five temperature event categories. The system used the index corresponding to the maximum probability as the event prediction label, enabling automatic identification of the temperature event type.

[0088] This method ultimately realizes an end-to-end temperature event recognition process with preprocessed temperature data and multidimensional feature combinations as input and event category labels as output. It significantly improves the ability to distinguish similar events (such as water baths and heating belts) and has good real-time and robustness.

[0089] Specific implementation method 2: Figure 2 、 Figures 7 to 9 This embodiment is described as an example of experimental verification using the distributed optical fiber temperature event recognition method based on the MFDS-GFNet network described in the first embodiment.

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

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

[0092] In this implementation, the performance of the intrusion event identification method is evaluated using three standard indicators: accuracy, precision, recall, and F1 score. They are calculated based on TP (true positive), FP (false positive), TN (true negative), and FN (false negative) in the confusion matrix, as shown below:

[0093]

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

[0095] like Figure 2 To more intuitively visualize the differences in key features between temperature events, this figure compares waterbath and heating zone events in three core characteristics: local standard deviation, temporal gradient, and spatial gradient. From the perspective of local standard deviation, waterbath events exhibit a uniform distribution of low values, indicating a consistent temperature field distribution; whereas heating zone events display higher local fluctuations, particularly at the boundaries. From the perspective of temporal gradient, waterbath events exhibit low-intensity, randomly distributed variations, while heating zone events exhibit distinct horizontal, banded regions of high gradient at certain time points, reflecting the staged nature of current regulation in the heating zone. From the perspective of spatial gradient, waterbath events exhibit significant gradients (bright streaks) only at the boundaries, with the gradient within the interior being almost zero; whereas heating zone events exhibit high gradients throughout the entire region, exhibiting an irregular pattern of bright spots. These differences in characteristics provide a reliable basis for the automatic identification of temperature events in distributed fiber optic sensing systems.

[0096] To verify the effectiveness of the feature selection method proposed in this implementation, a systematic evaluation of different feature combinations was conducted. The results showed that the selected three-feature combination (local standard deviation + temporal gradient + spatial gradient) achieved a recognition accuracy of 99.74%, significantly higher than that of single features and other feature combinations. In particular, the selected three-feature combination performed significantly better than the 98.85% accuracy achieved using all five features, demonstrating that feature complementarity is more important than the number of features.

[0097] like Figure 7 To verify the superiority of the proposed gated fusion mechanism, it was compared with three traditional fusion strategies: simple concatenation (Concat), addition fusion (Add), and attention fusion (Attention). Gate fusion demonstrated the best performance, achieving an overall accuracy of 99.74%, 0.56 percentage points higher than the second-best concatenation method (99.18%). In particular, for tropical events, gated fusion achieved 100% accuracy, while the other methods performed less well on such events (97.92% for concatenation, 96.86% for addition, and 97.38% for attention). This demonstrates that gated fusion can more effectively handle the feature relationships of complex events, especially for tropical events with their highly spatiotemporal variations.

[0098] In order to evaluate the contribution of each component in the MFDS-GFNet network model, a series of ablation experiments were conducted. The results show that the accuracy of the ResNet base model (no features, no SE module, no gated fusion) is 95.46%. Adding a single feature can increase the accuracy to 97.79%-98.80%, among which the spatial gradient feature has the best effect (98.80%). When the three key features are used in combination, the accuracy reaches 99.18%. Although the SE module is not significantly effective when added alone (96.88%), it plays a synergistic role in the complete model. The complete MFDS-GFNet model (including all components) achieved the best performance of 99.74%, proving the rationality of the design of each component.

[0099] To comprehensively evaluate the effectiveness of MFDS-GFNet, we compared it 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] Across all evaluation metrics, the MFDS-GFNet network model in this implementation demonstrated significant advantages, achieving a test accuracy of 99.74%, 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%). MFDS-GFNet also maintained high levels and balance in metrics such as precision, recall, and F1 score, demonstrating that the model not only has high overall accuracy but also performs stably across all event types, with no significant performance degradation for certain event types.

[0103] like Figure 8 As shown, the confusion matrix provides a clearer view of the different models' ability to distinguish five temperature events. Among the compared deep learning models, 1DCNN, the most basic time series model, exhibits significant confusion across multiple event types. In particular, the mutual misclassification rate between water bath and heating zone events reaches 8-10%, indicating that considering only temporal features is ineffective in distinguishing these two events, which have similar temperature ranges but different spatial distributions. Recognition performance gradually improves with increasing model complexity: 2DCNN reduces the misclassification rate between water bath and heating zone events to 4-5%. 2DCNN-BiLSTM further reduces the confusion between water bath and heating zone events (approximately 3%). GoogLeNet reduces the confusion rate to approximately 2%, and Transformer reduces it to approximately 1.5%. In contrast, MFDS-GFNet demonstrates superior discrimination. The confusion matrix shows that only a very small number of water bath samples (approximately 0.5%) are misclassified as heating zones, while recognition of all other event types reaches or approaches 100% accuracy.

[0104] like Figure 9 As shown in the figure, MFDS-GFNet converges the fastest among the six deep learning methods. Its validation accuracy rapidly rises to over 95% in the early stages of training (approximately 5 epochs), while the accuracy of other models at the same training stage is generally below 85%. In particular, MFDS-GFNet achieves over 98% accuracy in only about 15 epochs, while Transformer requires approximately 25 epochs, and GoogLeNet and 2DCNN-BiLSTM require over 30 epochs. The loss curve shows that MFDS-GFNet not only decreases the fastest but also has the smoothest curve, with significantly less fluctuation in the later stages of training than the other models. This demonstrates that MFDS-GFNet's training process is more stable and less susceptible to random fluctuations.

[0105] MFDS-GFNet's rapid convergence and training stability are attributed to its multi-feature fusion architecture, which enables the model to simultaneously utilize multiple complementary features, accelerating the search for optimal solutions. In particular, the adaptive nature of the gated fusion mechanism enables the model to dynamically adjust the importance of different features during training, quickly focusing on the most discriminative feature combinations. This feature is extremely valuable in practical applications, significantly reducing model training time and resource consumption.

[0106] In terms of computational efficiency, despite the introduction of multi-feature extraction and gated fusion mechanisms, MFDS-GFNet's parameter count (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), demonstrating that the proposed model strikes a good balance between complexity and performance. Regarding inference time, MFDS-GFNet processes each sample in 0.40 milliseconds, which is faster than simpler models such as 1DCNN (0.24 milliseconds) and 2DCNN (0.30 milliseconds), but lower than the Transformer (0.44 milliseconds), fully meeting the real-time processing requirements of practical applications.

[0107] Finally, the above experiments verified that the MFDS-GFNet network model of the present invention fully utilizes the spatiotemporal information of distributed optical fiber temperature data, which can better and faster identify various temperature events, especially for events with similar temperature characteristics (such as water baths and heating belts), with excellent distinguishing ability. Experimental results show that in the recognition tasks of five typical temperature events (background, water bath, heating belt, alcohol lamp, and freezing), MFDS-GFNet achieved a high accuracy rate of 99.74%, and achieved almost perfect recognition (accuracy rate of up to 100%) for the difficult-to-distinguish water bath and heating belt events. At the same time, the model exhibits faster convergence speed and more stable training process, as well as moderate computing resource requirements, and has high practical value.

[0108] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0109] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A distributed optical fiber temperature event recognition method based on the MFDS-GFNet network is characterized by: The implementation process of this method is: Step 1: Use the BOTDR system to collect the original temperature data of the distributed optical fiber, pre-process the original temperature data, and obtain the pre-processed temperature data; Step 2: Multi-feature extraction; Extract multiple features from the preprocessed temperature data and use feature importance Analyze and select the best feature set; Step 3: Construct the MFDS-GFNet network model, which consists of a dual-stream network module and a gated fusion module; During the training process, the temperature data preprocessed in step 1 and the optimal feature set extracted in step 2 are extracted through a dual-stream network module. The two feature streams are then integrated through a gated fusion module to achieve adaptive feature fusion. The classifier then outputs the probability distribution of temperature events. The training of the MFDS-GFNet network model is completed by optimizing the parameters of the MFDS-GFNet network model by minimizing the cross-entropy loss function and using the back-propagation algorithm. Step 4: During the test 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 automatic recognition of the temperature event type.

2. The distributed optical fiber temperature event identification method based on the MFDS-GFNet network according to claim 1, characterized in that: In step 1, the process of preprocessing the original temperature data is as follows: Step 11: Extract ROI from the original temperature data, calculate the temperature standard deviation of each spatial point in the time series, detect the temperature event area based on the distribution of the temperature standard deviation, and select the temperature event area as the ROI center; Step 1 and 2: Perform median filtering in both time and space dimensions to filter noise from the original temperature data; Step 1: Normalize the noise-filtered data to the range of [0, 1] to obtain the preprocessed temperature data X.

3. The distributed optical fiber temperature event identification method based on the MFDS-GFNet network according to claim 2, characterized in that: In step 1, 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, which is expressed as follows: in Where C i is the i-th cluster, C * is the most significant cluster, S(l) is the significance score at position l, and a ROI with a fixed width of w is extracted based on the center point c.

4. The distributed optical fiber temperature event identification method based on the MFDS-GFNet network according to claim 1, characterized in that: In step 2, the extracted multiple features include local statistical features, time gradient features, spatial gradient features, time second derivatives and temperature variation amplitude; The discrimination ability of each feature is quantified by calculating the ratio of the between-class variance to the within-class variance, and the optimal feature set is selected.

5. The distributed optical fiber temperature event identification method based on the MFDS-GFNet network according to claim 1, characterized in that: In step 3, the dual-stream network module is composed of a mainstream network and a tributary network; The mainstream network receives the temperature data preprocessed in step 1, extracts spatiotemporal features through a residual network structure, and obtains a mainstream feature vector; The tributary network receives the optimal feature set obtained in step 2 to generate a tributary feature vector; 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 the temperature event through the SoftMax classifier.

6. The distributed optical fiber temperature event identification method based on the MFDS-GFNet network according to claim 5, characterized in that: Both the mainstream network and the tributary network adopt the ResNet network structure; the ResNet network structure consists of three residual blocks, each residual block includes two residual units, a channel attention mechanism is integrated in each residual unit, and the feature map is converted into a fixed-length feature vector through global average pooling.

7. The distributed optical fiber temperature event identification method based on the MFDS-GFNet network according to claim 5, characterized in that: The gated fusion module concatenates the mainstream feature vector and the tributary feature vector in dimension to form a joint feature representation, generates a gating coefficient g through a fully connected layer and a Sigmoid activation function; finally, the gating coefficient is used to perform element-wise weighted fusion of the two feature vectors to obtain a fused feature vector.

8. The distributed optical fiber temperature event identification method based on the MFDS-GFNet network according to claim 7, characterized in that: The joint feature representation V concat , which can be expressed as follows: V concat =[V main ;V supp ]∈R 256 Where V main is the mainstream eigenvector, V supp is the tributary characteristic vector; The gating coefficient g is expressed as follows: g=σ(W g ·V concat +b g )∈R 128 Where W g and b g is the learnable parameter matrix and bias vector, σ is the Sigmoid activation function; Fusion feature vector V fused It can be expressed as follows: V fused =g⊙V main +(1-g)⊙V supp Where ⊙ represents the Hadamard product, which is an element-wise multiplication operation.

Images