Distributed optical fiber vibration wave classification method and system based on adaptive cross-domain feature fusion

Abstract

The invention relates to the field of distributed optical fiber sensing, and provides a distributed optical fiber vibration wave classification method and system based on adaptive cross-domain feature fusion. The method comprises the following steps: performing MFCC processing on a new phase sequence of a vibration signal to obtain an MFCC characteristic parameter of the vibration signal, and converting the MFCC characteristic parameter into a two-dimensional image; normalizing the new phase sequence of the vibration signal to obtain a one-dimensional sequence; based on the one-dimensional sequence, a one-dimensional convolutional neural network is adopted to obtain a time sequence feature vector; based on the two-dimensional image, adopting a feature extraction network to obtain an MFCC spectrogram feature vector; based on the time sequence feature vector and the MFCC spectrogram feature vector, a cross-domain attention mechanism and a difference perception gating unit are adopted to obtain a time sequence enhanced feature vector and an MFCC spectrogram enhanced feature vector; and carrying out adaptive weighted fusion on the time sequence enhanced feature vector and the MFCC spectrogram enhanced feature vector, and obtaining a classification result of the vibration signals based on a fusion vector.

Description

Technical Field

[0001] The present invention relates to the field of distributed optical fiber sensing technology, and in particular to a distributed optical fiber vibration wave classification method and system with adaptive cross-domain feature fusion. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Intelligent classification of fiber-optic vibration signals is a core requirement in the field of distributed fiber-optic sensing, but its accuracy and real-time performance remain challenging. With the increasing complexity of applications such as pipeline leak detection and perimeter security, traditional methods that rely on features in a single domain struggle to fully characterize the dynamic characteristics of vibration signals, limiting classification accuracy. For example, time-domain waveforms can capture transient impacts but ignore frequency-domain energy distribution, while spectral analysis is sensitive to periodic vibrations but susceptible to noise. Therefore, integrating time-domain and frequency-domain features is becoming increasingly important to improve classification performance.

[0004] However, achieving efficient and intelligent dual-domain feature fusion faces significant technical difficulties, and existing deep learning models still have many shortcomings in this regard.

[0005] First, at the level of cross-domain feature interaction, there is inherent heterogeneity between one-dimensional time series features and two-dimensional time-spectrogram features. They differ significantly in data dimension, numerical scale, information density, and the semantic level they carry. Direct or simple cross-domain attention mechanisms fail to fully account for the vast gap between different domains, potentially leading to inefficient information transfer and even the introduction of noise or irrelevant information, making it impossible to achieve true feature complementarity and synergistic enhancement. The lack of effective gating or regulation mechanisms to dynamically adapt to and manage this cross-domain information flow is a major bottleneck in current fusion methods.

[0006] Secondly, at the level of final feature weighted fusion, many existing methods employ simple feature concatenation post-processing or weighted summation based on static weights or basic attention mechanisms. These methods often overlook the dynamic changes in the information reliability and contribution of features from different domains under different input samples and interference conditions. For example, under strong narrowband noise interference, the reliability of frequency domain features may drop significantly. At this time, if they are still given a high fusion weight, it will contaminate the final decision. Existing fusion mechanisms lack the ability to conduct real-time reliability assessments of enhanced features in each domain and dynamically and adaptively adjust their contribution weights in the final fusion decision accordingly. This makes it difficult for the model to maintain optimal feature utilization efficiency and robustness in complex and changing environments, and key features are easily diluted by noisy channels or channels with low information content.

[0007] In addition, although some complex deep network structures can improve the fusion accuracy to a certain extent, their huge number of parameters and computational complexity make it difficult to deploy them on resource-constrained edge computing devices, and they cannot meet the urgent real-time requirements of many application scenarios. Summary of the Invention

[0008] To address the technical problems encountered in the aforementioned background technology, the present invention provides a distributed fiber oscilloscope classification method and system with adaptive cross-domain feature fusion. This method uses a one-dimensional time-domain waveform sequence and an MFCC (Mel-Frequency Cepstral Coefficient) feature spectrum as dual-channel inputs. A feature extraction network is used to capture the detailed characteristics of the signal's time-domain dynamic changes and frequency-domain energy distribution, respectively. A cross-domain attention mechanism is employed to achieve preliminary information exchange between the time-domain features and the MFCC spectrum features, allowing features from different domains to query and learn from each other. Furthermore, to further overcome the inherent differences between the time-domain and MFCC spectrum features and refine the information flow, a difference-aware gating unit is introduced. This gating unit dynamically adjusts the cross-domain features after the initial interaction, selectively enhancing beneficial complementary information and suppressing potential interference, thereby achieving more efficient and adaptable cross-domain feature enhancement. After the cross-domain interactive enhancement described above, the features of each domain are finally fused using an adaptive weighting method based on reliability and the attention mechanism. A reliability assessment module evaluates the relative reliability of each domain-enhanced feature in real time. The resulting reliability score is used to modulate and optimize the weights generated by the attention mechanism, ensuring that the final fusion weights more accurately reflect the true contribution and credibility of each domain in the current context, effectively suppressing the influence of noisy channels. While ensuring the effectiveness of feature extraction and the aforementioned advanced fusion mechanisms, the team focused on lightweight network structure design to meet the real-time inference requirements of distributed fiber optic vibration wave sensing systems on edge devices.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions: A first aspect of the present invention provides a distributed optical fiber vibration wave classification method with adaptive cross-domain feature fusion.

[0010] A distributed fiber oscillation wave classification method with adaptive cross-domain feature fusion, comprising: Obtaining the original phase sequence of the vibration signal collected by the distributed optical fiber vibration monitoring instrument, dividing the original phase sequence of the vibration signal, and obtaining a new phase sequence of the vibration signal; Perform MFCC processing on the new phase sequence of the vibration signal to obtain the MFCC feature parameters of the vibration signal and convert them into a two-dimensional image, that is, map them into the MFCC feature spectrum of the three-channel image; normalize the new phase sequence of the vibration signal to obtain a one-dimensional sequence; Based on the one-dimensional sequence, a one-dimensional convolutional neural network is used to obtain the time series feature vector; based on the two-dimensional image, a feature extraction network is used to obtain the MFCC spectrogram feature vector; based on the time series feature vector and the MFCC spectrogram feature vector, a cross-domain attention mechanism and a difference-aware gating unit are respectively used to obtain the time series enhanced feature vector and the MFCC spectrogram enhanced feature vector; the time series enhanced feature vector and the MFCC spectrogram enhanced feature vector are adaptively weighted fused to obtain the fused feature vector; Based on the fused feature vector, the classification result of the vibration signal is obtained.

[0011] Furthermore, the method of obtaining the MFCC spectrogram feature vector based on the two-dimensional image by using a feature extraction network includes: Convert the MFCC feature spectrum into a three-channel RGB image, input it into the initial convolution layer, and obtain the local feature vector; Based on the local feature vector, an inverted residual block group is used to obtain a fine-grained feature vector through multi-layer deep convolution and nonlinear transformation; Based on the fine-grained feature vector, the output convolution layer is used to expand the number of channels of the fine-grained feature vector while keeping the spatial resolution unchanged to obtain a high-dimensional feature vector; The global average pooling layer is used to compress the high-dimensional feature vector, and the MFCC spectrogram feature vector is obtained through the fully connected layer.

[0012] Furthermore, the method adopts a cross-domain attention mechanism and a difference perception gating unit based on the time series feature vector and the MFCC spectrogram feature vector to obtain a time series enhanced feature vector and an MFCC spectrogram enhanced feature vector respectively; the method includes: Both the time series feature vector and the MFCC spectrogram feature vector are normalized; Perform cross-domain attention calculation on the normalized time series feature vector and MFCC spectrogram feature vector respectively to obtain preliminary time series cross-domain interaction feature vector and MFCC spectrogram cross-domain interaction feature vector; The preliminary time series cross-domain interaction feature vector is input into the difference perception gating unit, and the first gating coefficient is calculated by combining the normalized time series feature vector; the preliminary MFCC spectrogram cross-domain interaction feature vector is input into the difference perception gating unit, and the second gating coefficient is calculated by combining the normalized MFCC spectrogram feature vector; the first gating coefficient is element-wise multiplied by the preliminary time series cross-domain interaction feature vector to obtain the final time series cross-domain interaction feature vector; the second gating coefficient is element-wise multiplied by the preliminary MFCC spectrogram cross-domain interaction feature vector to obtain the final MFCC spectrogram cross-domain interaction feature vector.

[0013] The final time series cross-domain interaction feature vector is added element by element to the normalized time series feature vector to obtain the first time series residual feature vector output by the first residual connection; the final MFCC spectrogram cross-domain interaction feature vector is added element by element to the normalized MFCC spectrogram feature vector to obtain the first MFCC residual feature vector output by the first residual connection; The normalized first time series residual feature vector and the first MFCC residual feature vector are input into the feedforward network to obtain the second time series residual feature vector and the second MFCC residual feature vector respectively; The second time series residual feature vector is added element by element to the normalized first time series residual feature vector to obtain the time series enhanced feature vector output by the second residual connection; the second MFCC residual feature vector is added element by element to the normalized first MFCC residual feature vector to obtain the MFCC enhanced feature vector output by the second residual connection.

[0014] Furthermore, the time series enhanced feature vector and the MFCC spectrogram enhanced feature vector are adaptively weighted fused to obtain a fused feature vector; the method includes: The time series enhanced feature vector and the MFCC spectrum enhanced feature vector are respectively input into a reliability assessment module after global average pooling, and a first reliability assessment factor and a second reliability assessment factor are output; A domain-shared context attention vector is introduced to obtain preliminary time series attention values ​​and preliminary MFCC spectrogram attention values. The preliminary time series attention values ​​are element-wise multiplied by the first reliability assessment factor to obtain the final time series attention value; the preliminary MFCC spectrogram attention values ​​are element-wise multiplied by the second reliability assessment factor to obtain the final MFCC spectrogram attention value; normalization is performed to obtain the weights of the time series enhanced feature vector and the weights of the MFCC spectrogram enhanced feature vector; According to the time series enhancement feature vector, the MFCC spectrogram enhancement feature vector, the weight of the time series enhancement feature vector and the weight of the MFCC spectrogram enhancement feature vector, a weighted fusion is performed to obtain a fused feature vector.

[0015] Furthermore, the method of obtaining a time series feature vector based on a one-dimensional sequence and using a one-dimensional convolutional neural network includes: Based on the one-dimensional sequence, multiple groups of convolution modules are used to extract multiple intermediate feature vectors; The extracted intermediate feature vectors are processed using flattening layers and fully connected layers to obtain time series feature vectors.

[0016] Furthermore, the new phase sequence of the vibration signal is subjected to MFCC processing to obtain MFCC feature parameters of the vibration signal and convert them into a two-dimensional image; the method includes: The new phase sequence of the vibration signal is sequentially subjected to pre-emphasis processing, frame processing and windowing processing; Performing a fast Fourier transform on the windowed signal to convert the windowed signal from the time domain to the frequency domain, obtaining the energy distribution of each frame of the vibration signal, and obtaining the energy spectrum of the vibration signal; Mel filter bank is used to perform Mel filtering on the energy spectrum of the vibration signal to obtain Mel filter energy; Perform logarithmic operation and discrete cosine transform on the Mel filter energy to obtain the MFCC feature parameters of the vibration signal; The first-order coefficients of the MFCC feature parameters of the vibration signal are discarded, and the logarithm of the energy of each frame is used as the new first-order coefficients, while the other 12-order coefficients remain unchanged; during the normalization process, the new first-order coefficients are normalized separately, and the other 12-order coefficients are normalized as a whole to obtain the updated MFCC feature parameters of the vibration signal; The updated MFCC feature parameters of the vibration signal are plotted as a spectrogram in the time-frequency domain and mapped into a two-dimensional image.

[0017] A second aspect of the present invention provides a distributed optical fiber vibration wave classification system with adaptive cross-domain feature fusion.

[0018] A distributed fiber oscillatory wave classification system with adaptive cross-domain feature fusion, comprising: A data acquisition module is configured to: acquire an original phase sequence of a vibration signal collected by a distributed optical fiber vibration monitoring instrument, segment the original phase sequence of the vibration signal, and obtain a new phase sequence of the vibration signal; The data processing module is configured to: perform MFCC processing on the new phase sequence of the vibration signal to obtain MFCC feature parameters of the vibration signal, and convert the parameters into a two-dimensional image, that is, map the MFCC feature spectrum of the three-channel image; normalize the new phase sequence of the vibration signal to obtain a one-dimensional sequence; The dual-channel network module is configured as follows: based on the one-dimensional sequence, a one-dimensional convolutional neural network is used to obtain a time series feature vector; based on the two-dimensional image, a feature extraction network is used to obtain an MFCC spectrogram feature vector; based on the time series feature vector and the MFCC spectrogram feature vector, a cross-domain attention mechanism and a difference-aware gating unit are used respectively to obtain a time series enhancement feature vector and an MFCC spectrogram enhancement feature vector; the time series enhancement feature vector and the MFCC spectrogram enhancement feature vector are adaptively weighted fused to obtain a fused feature vector; The classification module is configured to obtain a classification result of the vibration signal based on the fused feature vector.

[0019] A third aspect of the present invention provides a computer device, comprising: a processor adapted to execute a computer program; A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the distributed optical fiber vibration wave classification method with adaptive cross-domain feature fusion as described in the first aspect above.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which is suitable for being loaded by a processor and executing the steps of the distributed fiber vibration wave classification method with adaptive cross-domain feature fusion as described in the first aspect above.

[0021] A fifth aspect of the present invention provides a computer program product or computer program.

[0022] The present invention provides a computer program product or computer program, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the distributed fiber vibration wave classification method with adaptive cross-domain feature fusion as described in the first aspect above.

[0023] Compared with the prior art, the present invention has the following beneficial effects: The present invention fuses the dual-channel features of time-domain waveform and MFCC feature spectrum, combines lightweight 1D-CNN with improved MobileNetV2 feature extraction network to extract complementary information, and adopts a deeply optimized adaptive cross-domain feature fusion mechanism to ultimately achieve high-precision and low-latency distributed optical fiber vibration signal classification, providing reliable technical support for the fields of industrial monitoring and security.

[0024] The invention uses a unique cross-domain interaction stage, first using the basic attention mechanism to promote the preliminary mutual exploration and influence of feature information from different domains, so that the features of one domain can initially absorb contextual information from another domain. On this basis, the difference-aware gating unit is innovatively introduced to fine-tune the cross-domain information flow of this preliminary interaction. By intelligently perceiving and adapting to the inherent heterogeneity between the one-dimensional time domain sequence and the two-dimensional MFCC spectrogram features, it dynamically and selectively enhances beneficial complementary information while suppressing potential cross-domain interference or redundancy, thereby generating an enhanced feature representation that is purer, richer, more comprehensive, and more able to reflect the inherent correlation of the data than single domain information, simple feature splicing, or attention interaction without gated adjustment. Subsequently, an adaptive dual-domain feature fusion mechanism based on reliability and attention is adopted, which can dynamically calculate and assign the contribution weight of each enhanced domain feature to the final fusion result according to the specific characteristics of each input sample. This overcomes the limitations of traditional fusion methods where weights are fixed or require manual experience to set. The model can intelligently focus on domains in the current sample that are richer in information, higher in quality, or more relevant to the task objectives, thereby maximizing the utilization of effective information, suppressing the interference of noise or irrelevant information, and improving the accuracy of vibration signal classification.

[0025] By designing a lightweight 1D-CNN feature extraction network and an improved MobileNetV2 feature extraction network, the present invention reduces the total number of model parameters, facilitates deployment on edge devices, and meets the needs of real-time classification tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0027] Figure 1 This is a flow chart of a distributed optical fiber vibration wave classification method with adaptive cross-domain feature fusion according to an embodiment of the present invention; Figure 2 is a flow chart of a distributed optical fiber vibration wave classification method with adaptive cross-domain feature fusion according to another embodiment of the present invention; Figure 3 1 is a flow chart of MFCC feature parameter extraction shown in an embodiment of the present invention; Figure 4 is a structural diagram of a classification model shown in an embodiment of the present invention; Figure 5 1D-CNN feature extractor module according to an embodiment of the present invention; Figure 6 is a structural diagram of an adaptive cross-domain feature fusion module shown in an embodiment of the present invention; Figure 7 is a flow chart of calculating cross-domain attention weights shown in an embodiment of the present invention; Figure 8 1 is a structural diagram of a distributed optical fiber vibration wave classification system with adaptive cross-domain feature fusion according to an embodiment of the present invention; Figure 9 It is a structural diagram of a computer device shown in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0029] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0030] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0031] Figure 1 FIG. 1 is a flow chart of a distributed optical fiber vibration wave classification method with adaptive cross-domain feature fusion according to an embodiment of the present invention. Figure 1 As shown, the distributed fiber vibration wave classification method with adaptive cross-domain feature fusion described in this embodiment includes: Obtaining the original phase sequence of the vibration signal collected by the distributed optical fiber vibration monitoring instrument, dividing the original phase sequence of the vibration signal, and obtaining a new phase sequence of the vibration signal; Perform MFCC processing on the new phase sequence of the vibration signal to obtain the MFCC feature parameters of the vibration signal and convert them into a two-dimensional image, that is, map them into the MFCC feature spectrum of the three-channel image; normalize the new phase sequence of the vibration signal to obtain a one-dimensional sequence; Based on the one-dimensional sequence, a one-dimensional convolutional neural network is used to obtain the time series feature vector; based on the two-dimensional image, a feature extraction network is used to obtain the MFCC spectrogram feature vector; based on the time series feature vector and the MFCC spectrogram feature vector, a cross-domain attention mechanism and a difference-aware gating unit are respectively used to obtain the time series enhanced feature vector and the MFCC spectrogram enhanced feature vector; the time series enhanced feature vector and the MFCC spectrogram enhanced feature vector are adaptively weighted fused to obtain the fused feature vector; Based on the fused feature vector, the classification result of the vibration signal is obtained.

[0032] This paper introduces MFCC (Mel-Frequency Cepstral Coefficient) feature spectrograms into vibration signal classification, and uses a dual-channel feature extraction network to effectively extract time domain and frequency domain energy distribution features. Through adaptive cross-domain feature fusion, it enhances key information, suppresses noise, and effectively improves model recognition accuracy. The lightweight model meets the real-time inference requirements of edge devices.

[0033] The solution of this embodiment is described in detail below. Figure 2 As shown: Step S1: Data collection and preprocessing A distributed fiber-optic vibration monitoring instrument was used to collect raw phase sequences of vibration signals at a 1kHz pulse frequency, including those from ambient noise and three different interference events: a passing vehicle, manual excavation, and an operating excavator. Each raw phase sequence was 60,000 points long and spanned 60 seconds. The collected raw phase sequences were then preprocessed, including data segmentation, obtaining MFCC feature spectra, and data normalization.

[0034] First, data segmentation is performed on each original phase sequence. The original phase sequence is segmented into 2000 points as a standard length, and the corresponding time of the new sequence is 2 seconds.

[0035] MFCC feature parameters can be obtained by Figure 3 The extraction process is shown in the figure. First, the phase sequence is preprocessed, including pre-emphasis, framing, and windowing. Pre-emphasis can compensate for the high-frequency components of the signal, making the signal spectrum flatter.

[0036] Next, each vibration signal sequence is divided into frames, with a frame length of 40 milliseconds and a frame shift of 20 milliseconds.

[0037] Perform windowing on each frame of signal, and the window function used is Hamming window. The frame signal is recorded as , the sequence after windowing is ,in is the length of each frame signal sequence, is the total number of signal frames.

[0038] Next, the windowed signal is subjected to a fast Fourier transform (FFT) to obtain the transformed sequence ,in, Indicates the first spectral lines, Indicates the number of FFT points.

[0039] On this basis, the energy distribution of each frame of vibration signal can be obtained by squaring the spectrum amplitude of the signal sequence. .

[0040] Next, we use a Mel filter bank composed of triangular filters to perform Mel filtering on the energy spectrum. The order of the Mel filter is 22, which is the number of triangular filters in the Mel filter bank. The frequency response expression of the Mel filter is expressed as By performing Mel filtering on the energy spectrum, the Mel filter energy is obtained , the calculation formula is as follows: (1) in, represents the number of triangular filters, Indicates the A triangular filter.

[0041] Then, the Mel filter energy is logarithmized and discrete cosine transformed (DCT) to obtain the MFCC feature parameters of the vibration signal. The calculation formula is as follows: (2) in, Indicates the Frame signal MFCC parameters, is the number of required MFCC parameters, take The obtained MFCC feature parameters are the dimensions vector.

[0042] Finally, the first-order coefficients of the MFCC feature parameters are discarded, and the logarithm of the energy of each frame is used as the new first-order coefficients, and the remaining 12-order coefficients remain unchanged. During the normalization process, the new first-order coefficients are processed separately. Normalization, the remaining 12-order coefficients are overall Normalization.

[0043] The MFCC feature parameters corresponding to each vibration information are plotted as a spectrogram in the time-frequency domain and mapped into an RGB image to obtain a two-dimensional image sample.

[0044] At the same time, each vibration phase sequence is Normalize to obtain a one-dimensional data sample.

[0045] The present invention introduces the MFCC feature spectrum from speech signal processing technology into the task of optical fiber vibration signal classification. Compared with traditional time domain waveforms, the MFCC feature spectrum effectively suppresses environmental noise interference through frequency domain compression and focuses more on the characteristic representation of vibration events.

[0046] Step S2: Dataset construction Assign a unique ID to each sample and associate the one-dimensional and two-dimensional sample data. Create a CSV file that records the IDs, labels, and one-dimensional and two-dimensional sample paths for all samples. Based on the sample labels, use stratified sampling to divide the training, validation, and test sets into a ratio of 7:2:1.

[0047] Step S3: Building a dual-channel fiber optic vibration wave intelligent classification model The dual-channel fiber oscillation wave intelligent classification model includes a 1D-CNN feature extraction network, an improved MobileNetV2 feature extraction network, an adaptive cross-domain feature fusion module, and a Softmax classification layer. The overall structure of the model is as follows: Figure 4 shown.

[0048] Step S3.1: 1D-CNN feature extraction network This scheme proposes a 1D-CNN feature extraction network to capture the local features of one-dimensional vibration signal sequence. Figure 5 As shown in Figure 1, the 1D-CNN feature extraction network consists of 5 convolution groups, a Flatten layer and a fully connected layer.

[0049] The input of the network is a one-dimensional vibration signal sequence, which is subjected to feature extraction through multiple convolution modules. The size of the convolution kernel is 3, and the Relu activation function is used to enhance the nonlinear expression ability of the model. The padding of the convolution operation is set to 1. Maxpooling is used in the pooling layer to halve the size of the feature map. Dropout modules are added to the last two convolution modules to reduce the possibility of overfitting. A global average pooling operation is added at the end of the last convolution group to compress the one-dimensional feature map of each feature channel into a single value, thereby obtaining a 1024-dimensional feature representation. The Flatten layer is then used to ensure that the representation is a one-dimensional vector, and then a fully connected layer is used to map the dimension of the feature vector to 512 dimensions to facilitate fusion with the extracted two-dimensional spectrogram features.

[0050] Step S3.2: Improved MobileNetV2 feature extraction network The MobileNetV2 feature extraction network in this solution is an improvement on the existing MobileNetV2 network. It consists of an initial convolutional layer, six inverted residual module groups, an output convolutional layer, and a global average pooling layer. The six inverted residual module groups contain 1, 2, 3, 3, 2, and 2 inverted residual modules, respectively.

[0051] The input of this network is a three-channel RGB image converted from the MFCC spectrogram via pseudo color encoding.

[0052] The convolution kernel size of the initial convolution layer is , the number of filters is 32, the step size is 2, and the ReLU activation function is used to perform preliminary local feature extraction on the input MFCC spectrogram. The spatial resolution of the spectrogram is reduced by an operation with a step size of 2 to obtain the local feature vector.

[0053] The number of inverted residual block groups is set to 6, and their specific structure is the same as that of the standard MobileNetV2. The number of output channels after the six inverted residual block modules is 16, 24, 32, 64, 96, and 160, respectively. Through multiple layers of deep convolution and nonlinear transformations, the network is able to learn complex patterns and structural information from low-level to high-level in the MFCC spectrogram, generating fine-grained feature vectors.

[0054] The convolution kernel size of the output convolution layer is The number of filters is 512, and the ReLU activation function is used. While keeping the spatial resolution unchanged, the number of channels of the fine-grained feature vector is expanded to 512, and dimension enhancement and feature fusion are performed to obtain a high-dimensional feature vector.

[0055] The global average pooling layer compresses the final feature vector into a 1×1×512 feature vector.

[0056] The fully connected layer is 512-dimensional, which can further optimize the feature distribution, adapt to the requirements of multi-domain feature fusion, and obtain the MFCC spectrum feature vector.

[0057] Step S3.3: Adaptive cross-domain feature fusion module The adaptive cross-domain feature fusion module includes a cross-domain interaction module and an adaptive dual-domain feature fusion module based on reliability and attention. Its structure is as follows: Figure 6 shown.

[0058] The inputs of the cross-domain interaction module are the initial features in the time domain and the time-frequency domain, which come from the features of the one-dimensional vibration waveform sequence extracted by the 1D-CNN network and the features of the two-dimensional MFCC spectrogram extracted by the improved MobileNetV2 network. Indicates that or , are the one-dimensional time series feature vector and the two-dimensional MFCC spectrogram feature vector respectively.

[0059] First, the initial features are normalized for the first time. Input features of the domain Apply layer normalization to obtain normalized features , that is, the normalized time series feature vector and MFCC spectrogram feature vector.

[0060] The following is the calculation of cross-domain attention. The principle is as follows Figure 7 As shown. The cross-domain attention weight of the domain needs to be calculated Domain query vector and another domain The key vector Sum value vector , the calculation formula is as follows: (3) in, for The weight matrix of the domain, and for The weight matrix of the domain.

[0061] Then calculate the query by scaling the dot product and key The similarity score between them is calculated as follows: (4) in, Represents the vector dimension. Represents a query and key The similarity score between .

[0062] Apply the Softmax function to the score to get the normalized attention weight , the calculation formula is as follows: (5) in, Represents the query vector Key vector The attention weight of each element in is used to weight cross-domain features.

[0063] The attention weight Applied to a vector of values Perform weighted summation and get The domain absorbed Preliminary cross-domain interaction features based on attention for domain information , that is, the preliminary time series cross-domain interaction feature vector and the preliminary MFCC spectrogram cross-domain interaction feature vector, the calculation formula is as follows: (6) CDI is the abbreviation of Cross-Domain-Interaction.

[0064] The preliminary time series cross-domain interaction feature vector and the preliminary MFCC spectrogram cross-domain interaction feature vector are input into the difference perception gating unit respectively. The gating unit also considers the normalized features of the current domain And the preliminary cross-domain interaction features obtained through the attention mechanism mentioned above , and obtain a dynamic gating coefficient (Including the first gating coefficient and the second gating coefficient.) Gating coefficient The calculation formula is as follows: (7) in, Indicates that the feature and Concatenate along feature dimensions. and are the learnable weight matrix and bias vector of the linear layer in the gating unit, respectively. The sigmoid activation function constrains the output gating coefficient value to the interval (0,1), enabling it to effectively scale the information flow.

[0065] Then, using the learned gating coefficients Preliminary cross-domain interaction features Perform element-by-element multiplication to obtain the final cross-domain interaction features after gated refinement , that is, the final time series cross-domain interaction feature vector and the final MFCC spectrogram cross-domain interaction feature vector: (8) in, represents the element-wise product. It can more adaptively fuse effective information from another domain while suppressing potential noise or redundancy.

[0066] The cross-domain interaction features of the gated regulation calculated And the normalized features obtained Perform element-by-element addition to obtain the output of the first residual connection , that is, the first time series residual eigenvector and the first MFCC residual eigenvector. The calculation formula is as follows: (9) Then perform layer normalization on the added result to obtain the normalized features .

[0067] Will (The normalized first time series residual feature vector and the first MFCC residual feature vector) are input into a feedforward network The feedforward network consists of two linear layers and a nonlinear activation function between them, and has learnable parameters The features obtained by the feedforward network output are , that is, the second time series residual eigenvector and the second MFCC residual eigenvector. The calculation formula is as follows: (10) The output of the feedforward network Instead of input Add element by element to get the output of the second residual connection , that is, the time series enhanced feature vector and the MFCC spectrogram enhanced feature vector. The calculation formula is as follows: (11) Through the above steps, we can get the time series enhanced feature vector and MFCC spectrogram enhanced feature vector , which is the output of the second residual connection and .

[0068] In order to make the final fusion weight reflect the reliability or information contribution of different domains under specific input more precisely, the reliability evaluation mechanism is introduced into each enhanced domain feature through the following steps to obtain the reliability evaluation factor of the time series enhanced feature vector and reliability evaluation factor of MFCC spectrogram enhanced feature vector .

[0069] First, the time series enhanced feature vector and MFCC spectrogram enhanced feature vector The global average pooling operation is used to compress it into a fixed-size vector to capture its global context information. Subsequently, the two vectors are fed into a small multi-layer perceptron ( ), the perceptron consists of two fully connected layers and an activation function, which is used to learn and predict the original reliability score of the feature vector (Including the first reliability score and the second reliability score). The calculation formula is as follows: (12) in, Contains learnable weights and biases, Represents the time series enhanced feature vector Or MFCC spectrogram enhanced feature vector .

[0070] right Apply the sigmoid function to convert the original reliability score into a probabilistic reliability assessment factor between (0,1) , that is, the reliability assessment factor of the time series enhanced feature vector and reliability evaluation factor of MFCC spectrogram enhanced feature vector .

[0071] The input of the adaptive dual-domain feature fusion module based on reliability and attention is obtained in the cross-domain interaction stage. and And the reliability assessment factor obtained by the reliability assessment module and .

[0072] Next, we introduce a domain-shared contextual attention vector To obtain the initial attention value (including preliminary time series attention value and preliminary MFCC spectrogram attention value), the calculation process is as follows: (13) in, represents the shared attention vector, is the activation function, and represents the weights and biases of the linear layer, Represents the time series enhanced feature vector Or MFCC spectrogram enhanced feature vector .

[0073] The initial attention value Reliability assessment factor of the corresponding domain ( or ) performs element-wise multiplication to obtain the final attention value after reliability modulation .

[0074] Attention value for two domains Apply the Softmax function for normalization to obtain the final domain attention weight , the calculation formula is as follows: (14) in, is a scalar between 0 and 1, and . The larger the value, the The greater the contribution of the domain to the final fusion feature.

[0075] Use the calculated domain attention weights For the corresponding enhancement features Perform weighting and sum the weighted dual-domain features to obtain the final fusion feature vector , the calculation formula is as follows: (15) Step S4: Model training and testing Paired one-dimensional and two-dimensional data samples from the training set are fed into the dual-channel fiber oscillation wave intelligent classification model. Multi-class cross-entropy loss is used as the loss function to quantify the difference between the predicted and true labels. The average loss value for the current batch is calculated. The Adam optimization algorithm is then used to update the model parameters based on the calculated gradient information, gradually reducing the loss value. After training, the model parameter state corresponding to the best performance on the validation set is selected and saved as the final trained model for testing on the test set.

[0076] Step S5: Model deployment and intelligent classification After training the distributed fiber optic vibration wave intelligent classification model, it can intelligently classify vibration signals collected by the optical fiber. The trained intelligent classification model is exported to a format suitable for deployment and loaded into the monitoring system's operating environment. The one-dimensional vibration time series obtained from the distributed fiber optic vibration monitoring instrument is converted into an MFCC feature spectrum using the aforementioned steps. The two types of data are then input into the intelligent classification model to obtain the classification results for the vibration signal.

[0077] The present invention introduces MFCC characteristic spectra into the intelligent classification task of optical fiber vibration signals, effectively suppressing noise and highlighting the characteristics of different vibration signals.

[0078] The present invention adopts a cross-domain interaction module and an adaptive dual-domain feature fusion module based on reliability and attention to dynamically fuse dual-domain features, enabling the model to intelligently focus on domains with richer information, higher quality, or more relevant to the task objectives in the current sample, thereby maximizing the utilization of effective information and suppressing the interference of noise or irrelevant information.

[0079] Combination of the above Figure 1 、 Figure 2 The distributed optical fiber vibration wave classification method with adaptive cross-domain feature fusion provided by an embodiment of the present invention is introduced in detail. Next, the distributed optical fiber vibration wave classification system with adaptive cross-domain feature fusion provided by an embodiment of the present invention will be introduced with reference to the accompanying drawings.

[0080] Figure 8 This is a schematic diagram of the structure of a distributed optical fiber vibration wave classification system with adaptive cross-domain feature fusion according to an embodiment of the present invention. Figure 8 , the system of the present invention comprises: A data acquisition module is configured to: acquire an original phase sequence of a vibration signal collected by a distributed optical fiber vibration monitoring instrument, segment the original phase sequence of the vibration signal, and obtain a new phase sequence of the vibration signal; The data processing module is configured to: perform MFCC processing on the new phase sequence of the vibration signal to obtain MFCC feature parameters of the vibration signal, and convert the parameters into a two-dimensional image, that is, map the MFCC feature spectrum of the three-channel image; normalize the new phase sequence of the vibration signal to obtain a one-dimensional sequence; The dual-channel network module is configured as follows: based on the one-dimensional sequence, a one-dimensional convolutional neural network is used to obtain a time series feature vector; based on the two-dimensional image, a feature extraction network is used to obtain an MFCC spectrogram feature vector; based on the time series feature vector and the MFCC spectrogram feature vector, a cross-domain attention mechanism and a difference-aware gating unit are used respectively to obtain a time series enhancement feature vector and an MFCC spectrogram enhancement feature vector; the time series enhancement feature vector and the MFCC spectrogram enhancement feature vector are adaptively weighted fused to obtain a fused feature vector; The classification module is configured to obtain a classification result of the vibration signal based on the fused feature vector.

[0081] In some embodiments, the dual-channel network module is specifically configured as follows: converting the MFCC feature spectrum into a three-channel RGB image, inputting the initial convolution layer, and obtaining a local feature vector; based on the local feature vector, using an inverted residual block group, through multiple layers of deep convolution and nonlinear transformation, obtaining a fine-grained feature vector; based on the fine-grained feature vector, using an output convolution layer, while keeping the spatial resolution unchanged, expanding the number of channels of the fine-grained feature vector to obtain a high-dimensional feature vector; using a global average pooling layer to compress the high-dimensional feature vector, and obtaining the MFCC spectrum feature vector through a fully connected layer.

[0082] In some embodiments, the dual-channel network module is further configured to: normalize both the time series feature vector and the MFCC spectrogram feature vector; perform cross-domain attention calculation on the normalized time series feature vector and the MFCC spectrogram feature vector respectively to obtain a preliminary time series cross-domain interaction feature vector and an MFCC spectrogram cross-domain interaction feature vector; input the preliminary time series cross-domain interaction feature vector into the difference perception gating unit, and calculate the first gating coefficient in combination with the normalized time series feature vector; input the preliminary MFCC spectrogram cross-domain interaction feature vector into the difference perception gating unit, and calculate the second gating coefficient in combination with the normalized MFCC spectrogram feature vector; perform element-by-element multiplication of the first gating coefficient and the preliminary time series cross-domain interaction feature vector to obtain the final time series cross-domain interaction feature vector; perform element-by-element multiplication of the second gating coefficient and the preliminary MFCC spectrogram cross-domain interaction feature vector to obtain the final MFCC spectrogram cross-domain interaction feature vector. Quantity; the final time series cross-domain interaction feature vector is added element by element to the normalized time series feature vector to obtain the first time series residual feature vector output by the first residual connection; the final MFCC spectrogram cross-domain interaction feature vector is added element by element to the normalized MFCC spectrogram feature vector to obtain the first MFCC residual feature vector output by the first residual connection; the normalized first time series residual feature vector and the first MFCC residual feature vector are both input into the feedforward network to obtain the second time series residual feature vector and the second MFCC residual feature vector respectively; the second time series residual feature vector is added element by element to the normalized first time series residual feature vector to obtain the time series enhanced feature vector output by the second residual connection; the second MFCC residual feature vector is added element by element to the normalized first MFCC residual feature vector to obtain the MFCC enhanced feature vector output by the second residual connection.

[0083] In some embodiments, the dual-channel network module is further configured to: input the time series enhancement feature vector and the MFCC spectrogram enhancement feature vector into a reliability assessment module after global average pooling, and output a first reliability assessment factor and a second reliability assessment factor; introduce a domain-shared context attention vector to obtain a preliminary time series attention value and a preliminary MFCC spectrogram attention value, and perform element-by-element multiplication of the preliminary time series attention value and the first reliability assessment factor to obtain a final time series attention value; perform element-by-element multiplication of the preliminary MFCC spectrogram attention value and the second reliability assessment factor to obtain a final MFCC spectrogram attention value; perform normalization processing to obtain the weight of the time series enhancement feature vector and the weight of the MFCC spectrogram enhancement feature vector; perform weighted fusion according to the time series enhancement feature vector, the MFCC spectrogram enhancement feature vector, the weight of the time series enhancement feature vector and the weight of the MFCC spectrogram enhancement feature vector to obtain a fused feature vector.

[0084] In some embodiments, the dual-channel network module is further configured to: extract multiple intermediate feature vectors based on a one-dimensional sequence using multiple groups of convolution modules; and process the extracted intermediate feature vectors using a flattening layer and a fully connected layer to obtain a time series feature vector.

[0085] In some embodiments, the data processing module is specifically configured to: perform pre-emphasis processing, frame processing and windowing processing on the new phase sequence of the vibration signal in sequence; perform fast Fourier transform on the windowed signal to convert the windowed signal from the time domain to the frequency domain, obtain the energy distribution of each frame of the vibration signal, and obtain the energy spectrum of the vibration signal; use a Mel filter group to perform Mel filtering on the energy spectrum of the vibration signal to obtain Mel filter energy; perform logarithm operation and discrete cosine transform on the Mel filter energy to obtain the MFCC feature parameters of the vibration signal; discard the first-order coefficients of the MFCC feature parameters of the vibration signal, and use the logarithm of the energy of each frame as the new first-order coefficient, and the remaining 12-order coefficients remain unchanged; during normalization processing, the new first-order coefficients are normalized separately, and the remaining 12-order coefficients are normalized as a whole to obtain the updated MFCC feature parameters of the vibration signal; the updated MFCC feature parameters of the vibration signal are plotted as a spectrogram in the time-frequency domain and mapped into a two-dimensional image.

[0086] The present invention can adaptively and deeply fuse features from two heterogeneous data sources, one-dimensional time-domain series and two-dimensional time-spectrograms, fully leveraging their complementarity, overcoming their semantic gap, and effectively suppressing noise interference, while ensuring the lightweight model and meeting the needs of real-time inference of distributed fiber optic vibration wave sensing systems on edge devices.

[0087] The distributed optical fiber vibration wave classification system with adaptive cross-domain feature fusion according to the embodiment of the present invention may correspond to the method described in the embodiment of the present invention, and the above and other operations and / or functions of each module of the distributed optical fiber vibration wave classification system with adaptive cross-domain feature fusion are respectively to achieve Figure 1 For the sake of brevity, the corresponding processes of each method in are not repeated here.

[0088] See also Figure 9 The computer device shown in the figure includes a processor, a communication interface, and a computer-readable storage medium. The processor, communication interface, and computer-readable storage medium may be connected via a bus or other means. The communication interface is used to receive and transmit data. The computer-readable storage medium may be stored in the computer device's memory, the computer-readable storage medium storing a computer program comprising program instructions, and the processor executing the program instructions stored in the computer-readable storage medium. The processor (also known as a CPU (Central Processing Unit)) is the computing and control core of the computer device and is adapted to implement one or more instructions, specifically loading and executing one or more instructions to implement the corresponding steps in an embodiment of a distributed fiber vibration wave classification method with adaptive cross-domain feature fusion.

[0089] This embodiment provides a computer-readable storage medium (Memory). This computer-readable storage medium is a memory device within a computer device that is used to store programs and data. It should be understood that the computer-readable storage medium herein may include both built-in storage media within the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the processing system of the computer device. Furthermore, the storage space also stores one or more instructions suitable for being loaded and executed by the processor. These instructions may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium herein may be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device; alternatively, it may be at least one computer-readable storage medium located remotely from the processor.

[0090] In one embodiment, the computer-readable storage medium stores one or more instructions; the processor loads and executes the one or more instructions stored in the computer-readable storage medium to implement the corresponding steps in the above-mentioned embodiment of the distributed optical fiber vibration wave classification method with adaptive cross-domain feature fusion.

[0091] This embodiment provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the corresponding steps of the embodiment of the distributed fiber vibration classification method with adaptive cross-domain feature fusion.

[0092] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) containing computer-usable program code.

[0093] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0094] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0095] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0096] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0097] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A distributed fiber oscillation wave classification method based on adaptive cross-domain feature fusion, characterized by: include: Obtaining the original phase sequence of the vibration signal collected by the distributed optical fiber vibration monitoring instrument, dividing the original phase sequence of the vibration signal, and obtaining a new phase sequence of the vibration signal; Perform MFCC processing on the new phase sequence of the vibration signal to obtain the MFCC feature parameters of the vibration signal and convert them into a two-dimensional image, that is, map them into the MFCC feature spectrum of the three-channel image; normalize the new phase sequence of the vibration signal to obtain a one-dimensional sequence; Based on the one-dimensional sequence, a one-dimensional convolutional neural network is used to obtain the time series feature vector; based on the two-dimensional image, a feature extraction network is used to obtain the MFCC spectrogram feature vector; based on the time series feature vector and the MFCC spectrogram feature vector, a cross-domain attention mechanism and a difference-aware gating unit are respectively used to obtain the time series enhanced feature vector and the MFCC spectrogram enhanced feature vector; the time series enhanced feature vector and the MFCC spectrogram enhanced feature vector are adaptively weighted fused to obtain the fused feature vector; Based on the fused feature vector, the classification result of the vibration signal is obtained.

2. The distributed fiber oscillation wave classification method with adaptive cross-domain feature fusion according to claim 1 is characterized in that: The method of obtaining the MFCC spectrogram feature vector based on the two-dimensional image by using a feature extraction network includes: Convert the MFCC feature spectrum into a three-channel RGB image, input it into the initial convolution layer, and obtain the local feature vector; Based on the local feature vector, an inverted residual block group is used to obtain a fine-grained feature vector through multi-layer deep convolution and nonlinear transformation; Based on the fine-grained feature vector, the output convolution layer is used to expand the number of channels of the fine-grained feature vector while keeping the spatial resolution unchanged to obtain a high-dimensional feature vector; The global average pooling layer is used to compress the high-dimensional feature vector, and the MFCC spectrogram feature vector is obtained through the fully connected layer.

3. The distributed fiber oscillation wave classification method with adaptive cross-domain feature fusion according to claim 1 is characterized in that: The method adopts a cross-domain attention mechanism and a difference perception gating unit based on the time series feature vector and the MFCC spectrogram feature vector to obtain the time series enhanced feature vector and the MFCC spectrogram enhanced feature vector respectively; the method includes: Both the time series feature vector and the MFCC spectrogram feature vector are normalized; Perform cross-domain attention calculation on the normalized time series feature vector and MFCC spectrogram feature vector respectively to obtain preliminary time series cross-domain interaction feature vector and MFCC spectrogram cross-domain interaction feature vector; Input the preliminary time series cross-domain interaction feature vector into the difference perception gating unit, and combine it with the normalized time series feature vector to calculate the first gating coefficient; input the preliminary MFCC spectrogram cross-domain interaction feature vector into the difference perception gating unit, and combine it with the normalized MFCC spectrogram feature vector to calculate the second gating coefficient; perform element-by-element multiplication of the first gating coefficient and the preliminary time series cross-domain interaction feature vector to obtain the final time series cross-domain interaction feature vector; perform element-by-element multiplication of the second gating coefficient and the preliminary MFCC spectrogram cross-domain interaction feature vector to obtain the final MFCC spectrogram cross-domain interaction feature vector; The final time series cross-domain interaction feature vector is added element by element to the normalized time series feature vector to obtain the first time series residual feature vector output by the first residual connection; the final MFCC spectrogram cross-domain interaction feature vector is added element by element to the normalized MFCC spectrogram feature vector to obtain the first MFCC residual feature vector output by the first residual connection; The normalized first time series residual feature vector and the first MFCC residual feature vector are input into the feedforward network to obtain the second time series residual feature vector and the second MFCC residual feature vector respectively; The second time series residual feature vector is added element by element to the normalized first time series residual feature vector to obtain the time series enhanced feature vector output by the second residual connection; the second MFCC residual feature vector is added element by element to the normalized first MFCC residual feature vector to obtain the MFCC enhanced feature vector output by the second residual connection.

4. The distributed fiber oscillation wave classification method with adaptive cross-domain feature fusion according to claim 1 is characterized in that: The method of adaptively weighting and fusing the time series enhanced feature vector and the MFCC spectrogram enhanced feature vector to obtain a fused feature vector comprises: The time series enhanced feature vector and the MFCC spectrum enhanced feature vector are respectively input into a reliability assessment module after global average pooling, and a first reliability assessment factor and a second reliability assessment factor are output; A domain-shared context attention vector is introduced to obtain preliminary time series attention values ​​and preliminary MFCC spectrogram attention values. The preliminary time series attention values ​​are element-wise multiplied by the first reliability assessment factor to obtain the final time series attention value; the preliminary MFCC spectrogram attention values ​​are element-wise multiplied by the second reliability assessment factor to obtain the final MFCC spectrogram attention value; normalization is performed to obtain the weights of the time series enhanced feature vector and the weights of the MFCC spectrogram enhanced feature vector; According to the time series enhancement feature vector, the MFCC spectrogram enhancement feature vector, the weight of the time series enhancement feature vector and the weight of the MFCC spectrogram enhancement feature vector, a weighted fusion is performed to obtain a fused feature vector.

5. The distributed fiber oscillation wave classification method with adaptive cross-domain feature fusion according to claim 1 is characterized in that: The method is based on a one-dimensional sequence and uses a one-dimensional convolutional neural network to obtain a time series feature vector; the method includes: Based on the one-dimensional sequence, multiple groups of convolution modules are used to extract multiple intermediate feature vectors; The extracted intermediate feature vectors are processed using flattening layers and fully connected layers to obtain time series feature vectors.

6. The distributed fiber oscillation wave classification method with adaptive cross-domain feature fusion according to claim 1 is characterized in that: The method of performing MFCC processing on the new phase sequence of the vibration signal to obtain MFCC feature parameters of the vibration signal and converting them into a two-dimensional image includes: The new phase sequence of the vibration signal is sequentially subjected to pre-emphasis processing, frame processing and windowing processing; Performing a fast Fourier transform on the windowed signal to convert the windowed signal from the time domain to the frequency domain, obtaining the energy distribution of each frame of the vibration signal, and obtaining the energy spectrum of the vibration signal; Mel filter bank is used to perform Mel filtering on the energy spectrum of the vibration signal to obtain Mel filter energy; Perform logarithmic operation and discrete cosine transform on the Mel filter energy to obtain the MFCC feature parameters of the vibration signal; The first-order coefficients of the MFCC feature parameters of the vibration signal are discarded, and the logarithm of the energy of each frame is used as the new first-order coefficients, while the other 12-order coefficients remain unchanged; during the normalization process, the new first-order coefficients are normalized separately, and the other 12-order coefficients are normalized as a whole to obtain the updated MFCC feature parameters of the vibration signal; The updated MFCC feature parameters of the vibration signal are plotted as a spectrogram in the time-frequency domain and mapped into a two-dimensional image.

7. Distributed fiber oscillation wave classification system with adaptive cross-domain feature fusion, characterized by: include: A data acquisition module is configured to: acquire an original phase sequence of a vibration signal collected by a distributed optical fiber vibration monitoring instrument, segment the original phase sequence of the vibration signal, and obtain a new phase sequence of the vibration signal; The data processing module is configured to: perform MFCC processing on the new phase sequence of the vibration signal to obtain MFCC feature parameters of the vibration signal, and convert the parameters into a two-dimensional image, that is, map the MFCC feature spectrum of the three-channel image; normalize the new phase sequence of the vibration signal to obtain a one-dimensional sequence; The dual-channel network module is configured as follows: based on the one-dimensional sequence, a one-dimensional convolutional neural network is used to obtain a time series feature vector; based on the two-dimensional image, a feature extraction network is used to obtain an MFCC spectrogram feature vector; based on the time series feature vector and the MFCC spectrogram feature vector, a cross-domain attention mechanism and a difference-aware gating unit are used respectively to obtain a time series enhancement feature vector and an MFCC spectrogram enhancement feature vector; the time series enhancement feature vector and the MFCC spectrogram enhancement feature vector are adaptively weighted fused to obtain a fused feature vector; The classification module is configured to obtain a classification result of the vibration signal based on the fused feature vector.

8. A computer device, characterized in that: a processor adapted to execute a computer program; A computer-readable storage medium having a computer program stored therein, wherein when the computer program is executed by the processor, the computer program implements the steps of the distributed fiber vibration wave classification method with adaptive cross-domain feature fusion according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which is suitable for being loaded by a processor and executing the steps of the distributed fiber vibration wave classification method with adaptive cross-domain feature fusion according to any one of claims 1 to 6.

10. A computer program product, characterized in that The computer program product includes a computer program, and when the computer program is executed by a processor, the computer program implements the steps of the distributed optical fiber vibration wave classification method with adaptive cross-domain feature fusion according to any one of claims 1 to 6.

Images