Phi-OTDR vibration signal classification method based on multi-scale residual convolution and bidirectional long-short time memory network

Abstract

The invention discloses a phi-OTDR (Optical Time Domain Reflectometer) vibration signal classification method based on multi-scale residual convolution and a bidirectional long-short-term memory network, relates to the technical field of phase-sensitive optical time domain reflectometer vibration sensing, and solves the problems of insufficient feature extraction, single model fusion mechanism, poor system real-time performance and the like of an existing phi-OTDR vibration signal classification method. The method is realized through the steps of a multi-scale deep convolutional network, a bidirectional long-short term memory network, a cross attention fusion module, a learnable gating mechanism and the like. A cross attention fusion module is adopted, association between spatial features and time sequence features is dynamically established in a double-path network, feature interaction and information coupling are deepened, and the sensing ability of the model to key information is remarkably improved. Meanwhile, the fusion proportion of the multi-path features is adaptively adjusted according to the actual distribution of the input features, background noise interference is effectively suppressed, and the robustness and discrimination performance of the model are enhanced. Accurate identification and classification of external vibration signals are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of phase-sensitive optical time domain reflectometer (Φ-OTDR) vibration sensing, and particularly relates to a Φ-OTDR vibration signal classification method based on multi-scale residual convolution and bidirectional long short-term memory network. BACKGROUND

[0002] Φ-OTDR is a kind of distributed vibration monitoring technology based on optical fiber sensing, which is widely used in power communication line inspection, oil and gas pipeline monitoring, perimeter security and other scenes. This technology uses optical fiber itself as the sensing medium, and realizes accurate perception of external vibration events along the optical fiber path by detecting the slight changes in the phase of Rayleigh scattering light. Compared with traditional point sensors, Φ-OTDR has the advantages of wide coverage, strong distributed monitoring capability, and outstanding anti-electromagnetic interference capability, and thus has wide application value in the field of industrial and safety monitoring.

[0003] Among many application scenarios, the vibration monitoring of power communication lines is particularly typical. The lines are often widely distributed and have complex operating environments, and the collected vibration signals have the following characteristics: first, the vibration sources are diverse, including natural factors such as wind vibration, rain and snow, bird damage, and human factors such as external force construction, equipment knocking, and illegal damage; second, the noise background is complex, and there are strong electromagnetic interference and ground vibration in the signal, which are non-target components; third, different types of events (such as walking, digging, and rainfall) have significant differences in time sequence patterns and spectral characteristics, and traditional feature extraction methods are difficult to effectively capture and distinguish these information. The above characteristics result in strong nonlinearity, multi-scale and high randomness of the power communication line vibration data, which brings great challenges to signal classification. Traditional methods relying on manual feature extraction are difficult to fully capture these complex patterns, which can easily lead to high nuisance alarm rate (NAR) and insufficient classification accuracy, especially when multiple events exist simultaneously, the false alarm and missed alarm problems are particularly prominent.

[0004] In recent years, deep learning technology has been gradually applied to the automatic identification task of Φ-OTDR vibration signals. Convolutional neural network (CNN) has outstanding local feature extraction capability and performs outstandingly in capturing local events such as sudden impact, but is limited by the limited receptive field, and it is difficult to model the long time sequence dependence relationship in the power communication line. In contrast, the Transformer structure based on the self-attention mechanism can model long time series, but its parameter scale is large, which can easily cause high consumption of training resources and overfitting, especially in the case of limited data samples, the model generalization ability is difficult to guarantee.

[0005] To improve the modeling effect, some studies propose to use spatio-temporal joint modeling strategies, such as CNN-LSTM, CNN-Transformer and other combined structures. This kind of method often converts the original one-dimensional vibration signal into a two-dimensional image, uses two-dimensional convolution to extract spatial features, and then combines a time series model for classification. Although this "image" strategy can achieve high accuracy in experiments, as the monitoring distance or channel number increases, the data dimension grows exponentially, and the image conversion significantly increases the storage and computing overhead, and may dilute or even lose key information. In addition, most of these methods are simple module cascades, and the spatial and temporal feature fusion is not deep, making it difficult to fully exploit the coupling relationship between the time and space features in the power communication line vibration data.

[0006] More importantly, the image processing method has obvious limitations in actual engineering deployment. It relies on complex image conversion and two-dimensional neural networks, which often leads to long model loading time, high inference delay, and high energy consumption overhead and device overheating problems. On the front-end sensor nodes or edge computing devices of the power communication line, these problems are more prominent, severely limiting their application feasibility and scalability. Therefore, there is an urgent need for an efficient modeling method that does not require image conversion and can directly process the original one-dimensional vibration signal to adapt to the characteristics of the complex vibration data of the power communication line, while ensuring high recognition accuracy, reducing computational overhead and improving real-time performance and robustness.

[0007] To solve the problems of insufficient feature extraction, single model fusion mechanism and poor system real-time performance in existing Φ-OTDR vibration signal classification methods, the present application proposes a Φ-OTDR vibration signal classification method, which adopts a double-path modeling structure. One path extracts the spatial local features of the signal at different time scales through a multi-scale residual convolutional network, and the other path uses a bidirectional long short-term memory network (BiLSTM) to model the forward and backward dynamic evolution process of the vibration signal to obtain deep temporal features. On this basis, a cross-attention fusion module is introduced to dynamically establish the correlation between spatial and temporal features, improving the model's ability to perceive key information structures. Further combined with a learnable gating mechanism, the fusion ratio of each branch feature is adaptively regulated according to the actual distribution of the input features, effectively suppressing background noise interference and improving classification performance. The method directly acts on the original one-dimensional vibration signal without image conversion, avoiding the introduction of redundant computation and information loss, while ensuring high classification accuracy, significantly reducing inference delay and resource consumption, and having good real-time performance, lightweight characteristics and engineering deployment adaptability, suitable for efficient identification and intelligent analysis of various vibration events in the Φ-OTDR system. SUMMARY

[0008] The application provides a Φ-OTDR vibration signal classification method based on a multi-scale residual convolution and a bidirectional long short-term memory network to solve the problems of insufficient feature extraction, single model fusion mechanism and poor system real-time performance in existing Φ-OTDR vibration signal classification methods.

[0009] The Φ-OTDR vibration signal classification method based on the multi-scale residual convolution and the bidirectional long short-term memory network is specifically implemented through the following steps:

[0010] Step one, a Φ-OTDR vibration signal dataset is obtained, and a Z-Score normalization method is used to process the data to obtain processed vibration signal data;

[0011] Step two, a multi-scale deep convolutional neural network is constructed;

[0012] The vibration signal data obtained in step one is input into the multi-scale deep convolutional neural network, and a parallel multi-scale convolution kernel is used to extract spatial features F s of the vibration signal at different time scales;

[0013] Step three, a bidirectional long short-term memory network is used to extract time sequence features F t of the vibration signal data obtained in step one;

[0014] Step four, a cross-attention mechanism is introduced to fuse the spatial features extracted in step two and the time sequence features extracted in step three, that is, the time sequence features extracted by the bidirectional long short-term memory network in step three are taken as queries, the spatial features extracted by the multi-scale deep convolutional network in step two are taken as keys and values, the relationship between the two types of features is dynamically established through attention weights, and cross-attention fusion features F CA are obtained;

[0015] Step five, an adaptive gated feature fusion mechanism is used to adaptively adjust the relationship between the cross-attention fusion features and the original features; a fusion weight matrix is generated through feature splicing and a Sigmoid activation function, which is used to weight and integrate the original features and the cross-attention fusion features; finally, the fused feature vector is input into a full connection layer, and the prediction probability of each vibration signal category is calculated through a Softmax function.

[0016] The application has the following advantages:

[0017] 1. The application adopts Z-Score normalization and data enhancement methods to improve data quality and enhance the generalization ability of the model. Data enhancement includes adding Gaussian white noise and Poisson noise to enhance the robustness of the model to noise interference, and using sliding window and random cropping strategy to increase the diversity of data and improve the adaptability of the model to signals of different time lengths, thereby effectively reducing the risk of overfitting.

[0018] 2. The application adopts an innovative multi-scale deep convolution feature extraction module to extract local and global features of the vibration signal through parallel multi-scale convolution kernels, and combines a residual connection structure to effectively enhance the feature expression ability and improve the adaptability of the model to different types of vibration signals.

[0019] 3. The application adopts a bidirectional long short-term memory network that can simultaneously capture the forward and backward dependencies of the vibration signal, and compared to traditional unidirectional long short-term memory networks or recurrent neural networks, it can more comprehensively learn the time sequence features of the signal sequence, thereby improving the recognition ability of complex vibration patterns.

[0020] 4. The application adopts a cross-attention fusion module to dynamically establish the association between spatial features and time sequence features in the dual-path network, deepen feature interaction and information coupling, and significantly improve the model's perception ability of key information. At the same time, combined with a learnable gating mechanism, the fusion proportion of multi-path features is adaptively adjusted according to the actual distribution of input features, effectively suppressing background noise interference and enhancing the robustness and discrimination performance of the model. By introducing the cross-attention fusion module and the learnable gating mechanism, the high false positive rate of Φ-OTDR vibration signal classification is effectively reduced, and accurate recognition and classification of external vibration signals are realized.

[0021] 5. The application directly uses the original one-dimensional vibration signal as input, avoiding the information loss and computational redundancy caused by image or frequency spectrum in traditional methods, reducing the data preprocessing cost, preserving the key spatiotemporal characteristics of the original signal, and improving the overall performance and engineering practicability of the model.

[0022] 6. The method can effectively reduce the false positive rate of Φ-OTDR vibration signal classification and improve the classification accuracy, especially in the recognition of mining, intrusion, rainfall, walking and stress damage, etc. Compared with the prior art, it has wider application value in the fields of power transmission and perimeter security. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A flowchart of a Φ-OTDR vibration signal classification method based on multi-scale residual convolution and bidirectional long short-term memory network according to the application;

[0024] Figure 2A schematic diagram of the Z-Score normalization method described in the present application; wherein (a) is a schematic diagram of the original Φ-OTDR vibration signal waveform without normalization processing; (b) is a schematic diagram of the normalized waveform after Z-Score normalization processing;

[0025] Figure 3 A multi-scale residual convolutional network structure diagram described in the present application;

[0026] Figure 4 A residual convolutional module structure diagram described in the present application;

[0027] Figure 5 A bidirectional long short-term neural network structure diagram described in the present application;

[0028] Figure 6 A structure diagram of the cross-attention and adaptive gating fusion mechanism described in the present application;

[0029] Figure 7 A confusion matrix diagram described in the present application. DETAILED DESCRIPTION

[0030] The technical solutions of the present application will be described in detail below in combination with the accompanying drawings in the present application.

[0031] In combination with Figures 1 to 7 In this embodiment, a Φ-OTDR vibration signal classification method based on multi-scale residual convolution and bidirectional long short-term memory network is described, Figure 1 A Φ-OTDR vibration signal classification method flowchart provided in this embodiment, the method comprises:

[0032] Step 1, data preprocessing;

[0033] The Φ-OTDR vibration signal dataset is obtained, and is divided into a training set and a test set in a ratio of 8:2. The training set is subjected to data enhancement, including adding Gaussian white noise, Poisson noise and other interference, and increasing data diversity through sliding window and random cropping to improve the adaptability of the model to different types of vibration events. Subsequently, the Z-Score normalization method is used to process the data to eliminate the influence of data scale, improve data quality, enhance model generalization ability, reduce overfitting risk, improve model accuracy and stability.

[0034] In this embodiment, the Z-Score normalization method is used to normalize the vibration signal data, and the calculation formula is as follows:

[0035]

[0036] In the formula, X is the Φ-OTDR vibration signal to be processed; X 'is the Φ-OTDR vibration signal after Z-Score normalization processing; μ is the mean of the Φ-OTDR vibration signal to be processed, and σ is the variance of the Φ-OTDR vibration signal to be processed. This method can effectively convert the data distribution into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0037] As shown in Figure 2 , in the figure, (a) is the original Φ-OTDR vibration signal waveform of a typical rainfall event, and (b) is the normalized waveform after Z-Score normalization processing. By comparison, it can be seen that the normalization processing not only effectively eliminates the scale difference of the signal amplitude, but also highlights the essential feature structure of the signal, laying a foundation for subsequent feature extraction.

[0038] To further enhance the model's noise resistance and generalization performance, the embodiment designs a data enhancement strategy for the complex environment of power communication lines: in terms of noise injection, Gaussian white noise with a signal-to-noise ratio range of 15-30 dB and Poisson noise with an expectation λ of 0.05 are added to simulate various interference situations in the actual environment; in terms of time domain enhancement, a dynamic sliding window with a window length of 500 sampling points and a random cropping technique with a cropping ratio of 0.8-1.0 are used to effectively enhance the diversity of the data and improve the model's adaptability to signals of different time scales.

[0039] Through the combination of the above Z-Score normalization and data enhancement methods, the adaptability of different types of Φ-OTDR vibration signals can be effectively improved, the deviation of the data distribution can be reduced, and the recognition accuracy can be improved.

[0040] Step 2, construct a multi-scale deep convolutional neural network; used for extracting spatial features; extract local and global spatial features of vibration signals at different time scales through parallel multi-scale convolution kernels. The convolution layer adopts a deep residual connection mechanism to strengthen feature propagation and avoid gradient vanishing problems, thereby improving the robustness of the model in complex signal environments and enhancing the adaptability to multi-source vibration interference of power communication lines.

[0041] As shown in Figure 3 , in this embodiment, an efficient multi-scale deep convolutional neural network is constructed. This network introduces multi-scale residual convolution modules, cross-connection mechanisms, and cross-level connection mechanisms to effectively improve the recognition ability of vibration events. The network structure is described in detail as follows: let I be the Φ-OTDR vibration signal after preprocessing in step 1, the input first enters the initial convolution module Block0 to obtain the initial feature mapping:

[0042] I0=Block0(I)

[0043] Subsequently, the initial feature maps are sequentially passed through several residual convolutional modules Block1, Block2, ..., Block... n The output of each module is defined as:

[0044] I i =Block i (I i-1 ), i = 1, 2, ..., n

[0045] In each residual convolutional module Block1, Block2, ..., Block i ... Block n In the middle, the module structure is as follows Figure 4 As shown, it includes two parallel branches, each consisting of two convolutional layers, and introduces a cross-connection mechanism to extract local spatial features under different receptive fields.

[0046] First branch:

[0047]

[0048] Second branch:

[0049]

[0050] Where Conv k×1 This represents a convolution operation with a kernel size of k×1, where σ represents the ReLU activation function. The intermediate outputs of the two branches are cross-concatenated and then input into a 1×1 convolutional layer for feature fusion, and then combined with the input feature I. i-1 Perform residual connections to obtain the final output feature I of the i-th residual convolutional module. i :

[0051]

[0052] This structure can effectively alleviate the gradient vanishing problem and enhance the information transmission capability of deep networks.

[0053] To integrate semantic features from different depth layers, a cross-layer connection mechanism is introduced in the back-end design of the network to concatenate local and global features, constructing the final spatial feature vector (spatial feature) F. s ;

[0054] F s =Concat(I1,I2,…,I) i ,…,I n )

[0055] The spatial feature vector will be used as the key and value of the cross-attention mechanism in the subsequent steps, and the time sequence feature extracted by the BiLSTM (as the query) is deeply interacted and fused, so as to realize the space-time joint modeling of the Φ-OTDR vibration signal, and significantly improve the overall perception ability and classification performance of the model.

[0056] Step 3, extracting time sequence features by using a bidirectional long short-term memory neural network (BiLSTM); the BiLSTM can process the forward and backward information of the vibration signal at the same time, and comprehensively learn the time sequence features of the signal sequence. Compared with the traditional unidirectional long short-term memory network, the network can more accurately model the long-time dependence mode of the wind vibration, knocking, construction and other types of events, and improve the classification performance.

[0057] In the embodiment, the constructed BiLSTM network is as shown in the following figure: Figure 5 The BiLSTM network is composed of forward and backward LSTMs, and can capture the bidirectional dependence relationship of the vibration signal at the same time, thereby improving the modeling ability of the complex time sequence;

[0058] The calculation process of the BiLSTM network is as follows:

[0059]

[0060] In the formula, s t is the spatial feature vector input at the t-th time step, and respectively represent the hidden states calculated by the forward and backward LSTMs, and the output hidden state H t of the final BiLSTM is spliced by the forward and backward hidden states.

[0061] In the embodiment, the BiLSTM network controls the information flow through a gating mechanism, specifically including a forgetting gate, an input gate and an output gate, and the calculation formulas are as follows:

[0062] Forgetting gate: f t =σ(W f H t-1 ,s t +b f );

[0063] Input gate: i t =σ(W i [H t-1 ,s t +b i );

[0064] Output gate: o t =σ(W o [H t-1 ,s t +b o );

[0065] Cell state update: C t =f t ×C t-1 +i t ×tanh(W c [H t-1 ,s t ]+b c );

[0066] Final hidden state: H t =o t ×tanh(C t );

[0067] In the formula, W f W i W o W c b are the trainable weights of the BiLSTM f b i ,b o ,b c Here are the bias parameters, σ(·) is the Sigmoid activation function, tanh(·) is the hyperbolic tangent activation function, and C... t In cellular state, H t This represents the hidden state. By concatenating the hidden states from the forward and backward LSTMs, the complete temporal feature representation is finally obtained as F. t =[H1,H2,...,H T ], where T is the sequence length.

[0068] This temporal feature will be used as a query vector in subsequent steps to participate in the cross-attention mechanism and be deeply coupled and fused with the spatial features extracted by the multi-scale deep convolutional neural network, thereby realizing the spatiotemporal joint modeling of the Φ-OTDR vibration signal and significantly improving the overall perception capability and classification performance of the model.

[0069] Step 4: Feature fusion is achieved using a cross-attention fusion module. Temporal features extracted by BiLSTM are used as queries, and spatial features extracted by a multi-scale convolutional network are used as keys and values. The connection between the two types of features is dynamically established through attention weights, achieving deep coupling between spatial and temporal features. The fused features enhance the model's ability to capture key information and improve its performance in discriminating complex vibration events.

[0070] In this embodiment, the spatial feature F extracted in steps 2 and 3 is implemented using a cross-attention fusion module. s With time series features F t The deep interaction between them. The specific calculation process is as follows:

[0071] like Figure 6is defined as Q = W Q F t , the key matrix K = W K F s , and the value matrix V = W V F s , where W Q , W K , and W V are linear mapping matrices. The cross-attention calculation formula is as follows:

[0072]

[0073] In the formula, Z att is the feature mapping output by the cross-attention module; d k is the dimension of the key matrix.

[0074] The cross-attention mechanism dynamically weights the spatial features guided by the time sequence features, thereby generating a highly coupled fusion representation. Further through residual and normalization processing, the cross-attention fusion feature F CA is obtained:

[0075] F CA = LayerNorm (F t + Z att )

[0076] In the formula, LayerNorm is a normalization operation on the features in the channel dimension, which can stabilize training and speed up convergence; the module effectively enhances the complementary expression between spatial and time sequence features, highlights key patterns, and suppresses redundant interference.

[0077] Step 5, to achieve dynamic adjustment between the cross-attention fusion result and the original feature, an adaptive gated feature fusion mechanism is designed in this embodiment, which can further optimize the effect of fusion expression through feature adaptive weighting. Specifically, through feature splicing and Sigmoid activation function, a fusion weight matrix is generated for weighting and integrating original feature and fusion feature information, enhancing the robustness and generalization ability of the model in complex background. Finally, the fused feature vector is input into the fully connected layer, the prediction probability of each vibration event class is calculated through the improved Softmax function, and the performance of the model in accuracy, recall rate, precision rate, misjudgment rate, and F1 score and other indicators is evaluated using the test set.

[0078] First, F CA is spliced with the original feature F s +F t as input to calculate the fusion weight a:

[0079] a = σ (W g · [FCA ;F s +F t ]+b g )

[0080] where W g and b g are trainable gating parameters, σ is the Sigmoid activation function, and [·;·] represents the feature concatenation operation.

[0081] The final fusion feature Z is:

[0082] Z = α ⊙ F CA + (1-α) ⊙ (F s +F t )

[0083] where ⊙ is the element-wise multiplication operation. This design enables the model to dynamically adjust the contribution ratio of cross-attention features and original features according to the characteristics of the input signal, thereby preserving the essential features of the original signal while enhancing the expression of key discriminative information.

[0084] This mechanism enables the model to adaptively adjust the fusion ratio of cross-attention information and original feature information according to the input data distribution, enhancing the key pattern expression ability and effectively improving the robustness and discriminative performance of the system in complex scenarios.

[0085] Then, the fusion feature Z is passed through a fully connected layer and a Softmax activation function to calculate the probability distribution of each vibration event class:

[0086]

[0087] where W i ,b i are the fully connected layer parameters of the i-th class, C is the number of classes, and P(y = i|Z) represents the prediction probability of the i-th vibration event class. The system compares the probability values of the output sequence y1, y2, …, y6 and finally determines the class with the highest probability as the classification result of the vibration signal.

[0088] In this step, the test set divided in step 1 is used to verify the classification accuracy (Accuracy), recall (Recall), precision (Precision), false negative rate (NAR), and F1 score of the trained model. The calculation formulas are as follows:

[0089]

[0090] where TP i represents the number of samples of the i-th class that are correctly classified, FP i represents the number of samples of other classes that are incorrectly classified as the i-th class, and FN idenotes that the sample of the i-th class is misclassified as other classes. Since TN i It is difficult to define directly in the multi-classification task, and it is usually not considered as the main indicator, so it is usually not considered alone when calculating the overall accuracy. In order to balance the relationship between the precision and recall of the model, the F1 score is used to measure the performance of the classification model. The F1 score is the harmonic mean of precision and recall. The specific experimental results are shown in Figure 7 and Table 1, as shown in Figure 7 The performance indicators shown in Table 1 are as follows:

[0091] Table 1

[0092] Accuracy Recall Precision False dismissal rate F1 score Background noise 99.94％ 99.83％ 99.83％ 0.04％ 99.83％ Stress damage 99.97％ 99.80％ 100.00％ 0.00％ 99.90％ Digging 99.94％ 100.00％ 99.60％ 0.08％ 99.80％ Intrusion 99.94％ 99.78％ 99.78％ 0.04％ 99.78％ Rainfall 100.00％ 100.00％ 100.00％ 0.00％ 100.00％ Walking 99.84％ 99.39％ 99.59％ 0.08％ 99.49％ Mean 99.94％ 99.80％ 99.82％ 0.04％ 99.80％

[0093] As shown in Figure 7 , the prediction accuracy of each class is labeled below the corresponding label, and the accuracy of the six classes is 99.94%, 99.94%, 99.97%, 99.94%, 100.00% and 99.84%, respectively. The accuracy is between 99.84% and 100.00%, and the range is only 0.16%. The results are concentrated, indicating that the performance of the model on each class is stable and balanced. Table 1 shows the performance indicators of the Φ-OTDR vibration signal classification method based on multi-scale residual convolution and bidirectional long short-term memory network. The accuracy, recall, precision and F1 score of each class are high, and the misjudgment rate is low. The overall average accuracy of the model is 99.94%, the average recall is 99.82%, the average precision is 99.80%, the average F1 score is 99.80%, and the average misjudgment rate is only 0.04%.

[0094] The above results show that the model of the method can efficiently and accurately identify multiple vibration events on the power communication line, and performs excellently and stably in the multi-classification task.

[0095] The technical features of the above-described embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described, but as long as the combinations of the technical features do not contradict, they should be considered within the scope of the present disclosure.

[0096] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be construed as limiting the scope of the patent. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A vibration signal classification method based on multi-scale residual convolution and bidirectional long short-term memory network, characterized by: Step 1: Obtain the Φ-OTDR vibration signal dataset and process the data using the Z-Score normalization method to obtain the processed vibration signal data. Step 2: Construct a multi-scale deep convolutional neural network; The vibration signal data obtained in step one is input into a multi-scale deep convolutional neural network, and the spatial features F of the vibration signal at different time scales are extracted by parallel multi-scale convolutional kernels. s ; Step 3: Extract temporal features F from the vibration signal data obtained in Step 1 using a bidirectional long short-term memory network. t ; Step 4: Introduce a cross-attention mechanism to fuse the spatial features extracted in Step 2 and the temporal features extracted in Step 3. Specifically, use the temporal features extracted by the bidirectional long short-term memory network in Step 3 as the query, and the spatial features extracted by the multi-scale deep convolutional network in Step 2 as the key and value. Dynamically establish the relationship between the two types of features through attention weights to obtain the cross-attention fusion feature F. CA ; Step 5: Adaptive gating feature fusion mechanism is used to adaptively adjust the relationship between the cross-attention fusion features and the original features; By concatenating features and applying the Sigmoid activation function, a fusion weight matrix is ​​generated, which is used to weight and integrate the original features with the cross-attention fusion features. Finally, the fused feature vector is input into the fully connected layer, and the prediction probability of each vibration signal category is calculated by the Softmax function.

2. The Φ-OTDR vibration signal classification method based on multi-scale residual convolution and bidirectional long short-term memory network according to claim 1, characterized in that: In step one, the vibration signal data is normalized using the Z-Score normalization method, and the calculation formula is as follows: In the formula, X is the Φ-OTDR vibration signal to be processed; X ' σ represents the Z-Score normalized Φ-OTDR vibration signal; μ represents the mean of the Φ-OTDR vibration signal to be processed; and σ represents the variance of the Φ-OTDR vibration signal to be processed.

3. The Φ-OTDR vibration signal classification method based on multi-scale residual convolution and bidirectional long short-term memory network according to claim 1, characterized in that: In step two, the multi-scale deep convolutional neural network introduces a multi-scale residual convolution module, and extracts spatial features through cross-connection and cross-level connection mechanisms. Each residual convolutional module is configured to include two parallel branches, each consisting of two convolutional layers. A cross-connection mechanism is introduced to concatenate the intermediate outputs of the two branches, which are then fed into a 1×1 convolutional layer for feature fusion and combined with the input feature I. i-1 Perform residual connections to obtain output feature I i ; The features of the intermediate layer and the output layer are concatenated through a cross-level connection mechanism to construct the final spatial feature vector F. s .

4. The Φ-OTDR vibration signal classification method based on multi-scale residual convolution and bidirectional long short-term memory network according to claim 3, characterized in that: First branch: Second branch: In the formula, Conv k×1 For a convolution operation with a kernel size of k×1, σ represents the ReLU activation function. The intermediate outputs of the two branches are cross-concatenated and then input into a 1×1 convolutional layer for feature fusion, and then combined with the input feature I. i-1 Perform residual connections to obtain the final output of the i-th residual convolutional module; 5. The Φ-OTDR vibration signal classification method based on multi-scale residual convolution and bidirectional long short-term memory network according to claim 1, characterized in that: In step three, the bidirectional long short-term memory network consists of a forward LSTM and a backward LSTM, which simultaneously captures the bidirectional dependency of the vibration signal. The computation process of a bidirectional long short-term memory network is as follows: In the formula, s t Let be the spatial feature vector input at time step t. and These represent the hidden states computed by the forward and backward LSTMs, respectively, and the final output of the BiLSTM is the hidden state H. t The pieces are spliced ​​from the front and back hidden states.

6. The Φ-OTDR vibration signal classification method based on multi-scale residual convolution and bidirectional long short-term memory network according to claim 1, characterized in that: In step four, the cross-attention mechanism, guided by temporal features, dynamically weights spatial features to generate a highly coupled fusion representation; and through residual and normalization processing, it obtains the cross-attention fusion feature F. CA : F CA =LayerNorm(F t +Z att ) In the formula, LayerNorm is the normalization operation performed on the features along the channel dimension, and Z... att F is the feature map output by the cross-attention mechanism. t This represents the spatial characteristics of the output of a bidirectional long short-term memory network.

7. The Φ-OTDR vibration signal classification method based on multi-scale residual convolution and bidirectional long short-term memory network according to claim 1, characterized in that: In step five, an adaptive gating feature fusion mechanism is used to fuse the cross-attention fusion features F. CA With original feature F s +F t Using the concatenation as input, calculate the fusion weight α; α=σ(W g ·[F CA ;F s +F t ]+b g ) In the formula, W g and b g Here are the trainable gating parameters, and σ is the Sigmoid activation function; The final fused feature Z is: Z=α⊙F CA +(1-α)⊙(F s +F t ) In the formula, ⊙ represents element-by-element multiplication.

Images