An optical fiber acoustic-vibration dual-mode cable external damage positioning method and system

Abstract

The application relates to an optical fiber acoustic-vibration dual-mode cable external damage positioning method and system, and relates to the field of computer technology, which comprises a sensing optical fiber and a time sequence controller, acquires acoustic signals and vibration signals of a cable along a line, and performs time-space registration processing after acquisition to obtain dual-mode signals; characteristic processing is performed on the dual-mode signals to obtain vibration characteristic information and acoustic characteristic information; based on a dual-branch neural network model, vibration characteristic vectors and acoustic characteristic vectors are obtained; based on an attention mechanism, the vibration characteristic vectors and the acoustic characteristic vectors are subjected to weighted fusion processing to obtain acoustic-vibration characteristic vectors; based on the acoustic-vibration characteristic vectors, a cable external damage behavior recognition result is output through a classifier; based on the acoustic signals and the vibration signals, position information corresponding to the cable external damage behavior recognition result is determined, and an alarm is given through an alarm platform. The application has the effects of improving the accuracy of external damage behavior types and positions and the timeliness of external damage behavior alarms.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method and system for locating external damage to fiber optic acoustic-vibration dual-mode cables. Background Technology

[0002] Underground cables have become the preferred solution for urban power transmission construction due to their technical advantages such as short construction period, high space utilization, and strong resistance to environmental interference. However, with the acceleration of urbanization, underground power cables are frequently damaged by external factors such as construction excavation and mechanical collisions, which seriously affects the safe operation of the power grid.

[0003] Among related technologies, monitoring methods for external cable damage include: single vibration sensor monitoring and distributed fiber optic sensing systems. Single vibration sensor monitoring technology mainly involves burying vibration sensing devices such as accelerometers near the cable to collect mechanical vibration signals propagating from the ground. When the vibration amplitude exceeds a preset threshold, it is determined as an external damage event and an alarm is triggered. Distributed fiber optic sensing systems primarily utilize the communication optical fiber attached to the cable as the sensing medium. Through distributed acoustic wave sensing technology, probe light pulses are injected into the fiber, and the external damage behavior is determined by analyzing the phase or intensity changes of Rayleigh scattered light, thus achieving monitoring of the acoustic or vibration field along the cable.

[0004] Regarding the aforementioned technologies, while a single vibration sensor can detect ground vibrations, it struggles to distinguish between ambient vibrations (e.g., vehicle traffic and machinery operation) and actual external damage events, resulting in a high false alarm rate. Furthermore, sensors can only output single vibration intensity information, failing to effectively identify the behavioral characteristics of different construction equipment (e.g., excavators and hydraulic breakers), thus hindering accurate early warning. Distributed fiber optic acoustic sensing systems employ single-mode signal acquisition (e.g., acquiring only acoustic signals or only vibration signals), leading to inaccurate identification of specific external damage behaviors in complex noise environments and preventing timely early warning. There is still room for improvement. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, such as low accuracy of single detection and identification, difficulty in determining specific external damage behavior, large spatial positioning error, and delayed alarm response, this application provides a fiber optic acoustic-vibration dual-mode cable external damage positioning method and system.

[0006] Firstly, this application provides a method for locating external damage to a fiber optic acoustic-vibration dual-mode cable, employing the following technical solution: Based on a single sensing fiber and a timing controller, acoustic and vibration signals along the cable are collected. These signals are then spatiotemporally registered to obtain dual-mode signals. Feature processing is performed on the dual-mode signals to obtain vibration and acoustic feature information. Based on a dual-branch neural network model, feature extraction is performed on the vibration and acoustic feature information to obtain vibration and acoustic feature vectors. Using an attention mechanism, the vibration and acoustic feature vectors are weighted and fused to obtain an acoustic-vibration feature vector. Based on this feature vector, a classifier outputs the cable external damage identification result. Based on the acoustic and vibration signals, the location information corresponding to the identified cable external damage is determined, and the identification result and location information are sent to an alarm platform for alarm activation.

[0007] Secondly, this application provides a fiber optic acoustic-vibration dual-mode cable external damage location method system, which adopts the following technical solution: The acquisition module is used to acquire acoustic and vibration signals; the memory is used to store the program of the above-mentioned fiber optic acoustic-vibration dual-mode cable external damage location method; the processor is used to load and execute the program in the memory and implement the above-mentioned fiber optic acoustic-vibration dual-mode cable external damage location method.

[0008] The above-described embodiments of this disclosure have the following beneficial effects: the fiber optic acoustic-vibration dual-mode cable external damage location method of some embodiments of this disclosure improves the accuracy of external damage behavior type and location, and the timeliness of external damage behavior alarm. Specifically, the reason for the difficulty in determining specific external damage behavior, large spatial positioning error, and delayed alarm response is that the sensor can only output single vibration intensity information and cannot effectively identify the behavioral characteristics of different construction equipment (e.g., excavators and hydraulic breakers), thus failing to provide accurate early warning. In complex noise environments, the identification of specific external damage behaviors is inaccurate, and timely early warning cannot be provided. Based on this, the fiber optic acoustic-vibration dual-mode cable external damage location method of some embodiments of this disclosure firstly collects acoustic wave signals and vibration signals along the cable based on a sensing fiber and a timing controller, and performs spatiotemporal registration processing on the aforementioned acoustic wave signals and vibration signals to obtain dual-mode signals. Thus, synchronous acquisition of acoustic wave and vibration signals at the same physical location and time is achieved, providing a data foundation for subsequent fusion of multi-mode information and avoiding errors in cable external damage behavior identification caused by signal asynchrony. Secondly, feature processing is performed on the aforementioned dual-modal signals to obtain vibration and acoustic feature information. This extracts physical features reflecting mechanical impact characteristics from the vibration signal and cepstral and spectral features reflecting the acoustic signature of the equipment from the acoustic signal, transforming the original signal into interpretable feature information and providing high-quality input for neural network learning and inference. Then, based on a dual-branch neural network model, feature extraction is performed on the aforementioned vibration and acoustic feature information to obtain vibration and acoustic feature vectors. Thus, the vibration branch of the dual-branch neural network model is used to explore the impact period and energy attenuation law of the vibration signal, while the acoustic branch characterizes the acoustic signature details of different construction equipment, enabling the model to learn discriminative feature information from each mode. Furthermore, based on an attention mechanism, the aforementioned vibration and acoustic feature vectors are weighted and fused to obtain an acoustic-vibration feature vector. This achieves adaptive fusion of the two modal information, dynamically adjusting the contribution of vibration and acoustic waves according to signal quality. In complex noise environments, it can highlight effective modes and suppress interfering modes, improving the robustness and discriminative ability of the fused features. Next, based on the aforementioned acoustic and vibration feature vectors, a classifier outputs the cable external damage behavior identification results. This allows for accurate differentiation of different external damage behaviors, overcoming the deficiency of simply triggering an alarm without recognizing the actual damage. Finally, based on the aforementioned acoustic and vibration signals, the location information corresponding to the cable external damage behavior identification results is determined, and both the identification results and location information are sent to the alarm platform for alerting.Therefore, while identifying the type of damage, precise spatial location is achieved. The alarm information, including the type of external damage and its precise coordinates, is pushed to the maintenance terminal in real time via an IoT platform, solving the problems of large positioning errors and delayed alarms, and supporting rapid response and precise handling of external damage incidents. In summary, when monitoring cable external damage, the accuracy of the damage type and location is improved, as well as the timeliness of alarms, supporting rapid decision-making and on-site handling by maintenance personnel, and ensuring the stable operation of the power grid. Attached Figure Description

[0009] Figure 1 This is a flowchart of some embodiments of the fiber optic acoustic-vibration dual-mode cable external damage location method according to the present disclosure. Detailed Implementation

[0010] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0011] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0012] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0013] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0014] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0015] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0016] refer to Figure 1 The flowchart 100 illustrates some embodiments of the fiber optic acoustic-vibration dual-mode cable external damage location method according to this disclosure. The fiber optic acoustic-vibration dual-mode cable external damage location method includes the following steps: Step 101: Based on a sensing fiber and a timing controller, acquire acoustic and vibration signals along the cable, and perform spatiotemporal registration processing on the acoustic and vibration signals to obtain dual-mode signals.

[0017] In some embodiments, the execution entity (e.g., an electronic device) of the above-described fiber optic dual-mode cable damage location method can be hardware or software. When the computing device is hardware, it can be implemented as a distributed cluster of multiple servers or terminal devices, or as a single server or a single terminal device. When the computing device is software, it can be installed in the hardware devices listed above. It can be implemented as multiple software programs or software modules to provide distributed services, or as a single software program or software module. No specific limitations are made here.

[0018] In some embodiments, the aforementioned execution entity can acquire acoustic and vibration signals along the cable based on a sensing optical fiber and a timing controller, and perform spatiotemporal registration processing on the acoustic and vibration signals to obtain a dual-mode signal. The dual-mode signal can be a dataset containing both acoustic and vibration signals. The acoustic signal can be an electrical signal representing changes in acoustic waves. The vibration signal can be an electrical signal representing changes in vibration.

[0019] In some optional implementations of certain embodiments, the aforementioned execution entity can acquire acoustic and vibration signals along the cable based on a sensing optical fiber and a timing controller, and perform spatiotemporal registration processing on the acoustic and vibration signals to obtain a dual-mode signal, which may include the following steps: The first step involves coupling a first probe light and a second probe light into a sensing optical fiber using wavelength division multiplexing (WDM). The first probe light is used to excite and acquire acoustic wave signals, while the second probe light is used to excite and acquire vibration signals. The first probe light can be a laser pulse with a wavelength of 1550 nm, and the second probe light can be a laser pulse with a wavelength of 1625 nm.

[0020] As an example, firstly, an integrated dual-wavelength laser source can be used to generate a first probe beam (e.g., 1550 nm) and a second probe beam (e.g., 1625 nm). Then, the two beams can be coupled into a sensing fiber (e.g., G.652.D single-mode communication fiber) using a wavelength division multiplexer. This fiber can serve as both a transmission medium and a distributed sensor. Within the fiber, the light pulses of the first probe beam are used for phase-sensitive optical time-domain reflectometry (φ-OTDR) detection. External acoustic waves (pressure waves) acting on the fiber cause micrometer-level elastic deformation of the fiber's refractive index, thereby modulating the phase of the backscattered Rayleigh light. Demodulating this phase change allows the acoustic signal to be recovered. The second probe beam is used for distributed acoustic sensing (DAS) detection. External vibrations cause axial strain in the fiber; by detecting changes in the intensity of the backscattered Rayleigh light, the vibration signal can be demodulated. After the backscattered light of the two wavelengths returns, it is separated by a wave demultiplexer and sent to the first probe optical channel detector (high-sensitivity balanced photodetector) and the second probe optical channel detector (high-speed photodiode) for photoelectric conversion.

[0021] The second step involves synchronously sampling the optical signal along the cable using a dual-channel synchronous data acquisition card and the aforementioned sensing fiber, obtaining acoustic and vibration signals. The acoustic signal can be an electrical signal obtained by demodulating the phase change of the Rayleigh scattered light after the first probe light is emitted. The vibration signal can be an electrical signal obtained by detecting the intensity change of the Rayleigh scattered light after the second probe light is emitted.

[0022] As an example, firstly, the signals received by the two detectors in a single sensing fiber are connected to the two analog-to-digital converters (ADCs) of a dual-channel synchronous data acquisition card (DAQ). The sampling clock of this dual-channel synchronous data acquisition card is directly locked by the master clock provided by a timing controller (e.g., an FPGA), fundamentally ensuring strict synchronization of the sampling times of the two signals. Then, the two returned optical signals can be synchronously sampled to obtain an acoustic signal (acquired by the first probe light) and a vibration signal (acquired by the second probe light).

[0023] The third step involves determining the hardware timestamp information corresponding to the acoustic wave signal and the vibration signal based on the synchronous trigger clock generated by the timing controller, thereby obtaining the acoustic wave signal data packet and the vibration signal data packet. The hardware timestamp information can be the time stamp information generated by the timer at the instant the laser pulse is emitted.

[0024] As an example, a high-performance field-programmable gate array (FPGA) can be used as the core timing controller. A fully synchronized trigger clock is generated to simultaneously drive the first probe light and its laser pulse modulation circuit, ensuring that both optical pulses are injected into the optical fiber at the same absolute time point, generating a hardware timestamp (nanosecond level). Then, the hardware timestamp can be written into the fixed header of the data packet to obtain the acoustic signal data packet and the vibration signal data packet.

[0025] The fourth step involves extracting hardware timestamp information from the acoustic signal data packet and vibration signal data packet, and performing time-domain shift compensation on the sampling sequence of the acoustic channel based on a fixed transmission delay difference, to obtain an acoustic-vibration aligned signal. Here, the fixed transmission delay difference is the transmission delay difference between the acoustic signal and the vibration signal. The aligned signal can be an acoustic signal and a vibration signal aligned on the time axis.

[0026] As an example, firstly, after the system is initially installed or reset, a standard transient excitation (e.g., a short tap from a standard tapper) is applied near the fiber optic cable (approximately 10 meters from the sensor host). This standard transient excitation can simultaneously generate both acoustic and vibration signals detectable in two modes. Then, both signals can be acquired, and hardware timestamp information can be extracted from the acoustic and vibration signal data packets. Next, the peak arrival time of the signal is detected as... (Vibration channel) and (Sound wave channel). Calculate the fixed delay difference: Among them, the above It can be the fixed transmission delay between the sound wave channel and the vibration channel.

[0027] During real-time acquisition, for the continuously acquired data stream (sampling sequence) of the acoustic channel, overall displacement compensation is performed in the time domain. The compensation amount can be... In practice, time-domain shift compensation can be performed using interpolation algorithms.

[0028] Fifth, for each acoustic-vibration alignment signal point in the above-mentioned acoustic-vibration alignment signal, based on the hardware timestamp information and the spatial positioning information of the fiber optic sensing system, a spatiotemporal tag is generated for each acoustic-vibration alignment signal point, resulting in a spatiotemporal tag sequence. In practice, spatiotemporal tags can be generated based on the hardware timestamp and spatial positioning information corresponding to the signal point. For example, the spatiotemporal tag could be... Among them, the above It can be a spatial location. (The above) This could be a sampling time point. (The above...) This can be the sampling interval. (The above...) It can be a hardware timestamp.

[0029] The sixth step is to encapsulate the above-mentioned spatiotemporal tag sequence and the above-mentioned acoustic-vibration alignment signal to obtain a dual-mode signal.

[0030] As an example, the aligned and labeled acoustic-vibration aligned signals (vibration signals and sound wave signals) are packaged by time windows (e.g., 1 second) with the data packet structure in JSON format to obtain a bimodal signal.

[0031] Step 102: Perform feature processing on the dual-mode signal to obtain vibration feature information and acoustic wave feature information.

[0032] In some embodiments, the execution entity can perform feature processing on the dual-modal signal to obtain vibration feature information and acoustic wave feature information. The vibration feature information can be feature information possessing both time-domain and frequency-domain characteristics of the vibration signal. The acoustic wave feature information can be feature information possessing both time-series characteristics and energy distribution characteristics of the acoustic wave signal.

[0033] In some optional implementations of certain embodiments, the execution entity may perform feature processing on the dual-modal signal to obtain vibration feature information and acoustic wave feature information, which may include the following steps: The first step is to perform multi-scale decomposition on the vibration signal in the aforementioned dual-mode signal to obtain vibration signal components. This multi-scale decomposition can be a method of decomposing the signal at different frequency scales. The vibration signal components can be sub-signals corresponding to different frequency ranges after multi-scale decomposition. As an example, wavelet transform can be used to decompose the vibration signal. Specifically, the db4 wavelet basis function can be used for the wavelet transform.

[0034] The second step is to denoise the aforementioned vibration signal components to obtain a denoised vibration signal. This denoised vibration signal can be a vibration signal that has undergone noise suppression. In practice, a heuristic thresholding criterion can be used to threshold the wavelet coefficients at each level. This heuristic thresholding criterion can automatically determine the threshold based on the entropy characteristics of the signal. Then, the thresholded wavelet coefficients at each level are reconstructed to obtain the denoised vibration signal.

[0035] The third step involves performing time-domain analysis on the denoised vibration signal to obtain time-domain characteristic information, including: root mean square (RMS), peak factor, and zero-crossing rate. The RMS can be a statistical measure characterizing the average energy of the signal. The peak factor can be an indicator reflecting the signal's impulse characteristics. The zero-crossing rate is the number of times the signal crosses zero level per unit time. The zero-crossing rate can also reflect changes in the signal's frequency components.

[0036] The fourth step involves performing frequency domain transformation on the denoised vibration signal to obtain frequency domain feature information. This frequency domain feature information includes: spectral centroid, spectral width, and subband energy ratio. The frequency domain transformation can be a method of converting the signal from the time domain to the frequency domain. The spectral centroid can be a statistical quantity characterizing the center frequency position of the spectrum. The formula for calculating the spectral centroid can be... The above. It can be a frequency value. (The above) It can be the first The frequency amplitude corresponding to each frequency point. The above. It can be an index value representing a frequency point. (The above...) This could be the total number of frequency points. The aforementioned spectral width can be a statistical measure characterizing the extent of spectral spread. The formula for calculating the aforementioned spectral width can be... The above. It can be a frequency value. (The above) It can be the first The frequency amplitude corresponding to each frequency point. The above. It can be an index value representing a frequency point. (The above...) This could be the total number of frequency points. (The above...) This could be the spectral centroid value. The aforementioned sub-band energy ratio can be the proportion of energy in each sub-band to the total energy after dividing the frequency band into multiple sub-bands. For example, performing a Fast Fourier Transform on a denoised vibration signal and dividing the frequency band into three sub-bands—low frequency (0-200Hz), mid frequency (200-1000Hz), and high frequency (>1000Hz)—can calculate the proportion of energy in each sub-band to the total energy, and thus calculate the spectral centroid and spectral width.

[0037] The fifth step is to integrate the aforementioned time-domain and frequency-domain feature information to obtain the vibration feature information. In practice, this integration can be achieved using vector concatenation.

[0038] The sixth step is to perform acoustic feature processing on the acoustic signal in the above dual-mode signal to obtain acoustic feature information.

[0039] In some optional implementations of certain embodiments, the execution entity may perform acoustic feature processing on the acoustic signal in the dual-modal signal to obtain acoustic feature information, which may include the following steps: The first step involves performing spectral flattening and framing / windowing on the aforementioned acoustic signal to obtain an acoustic frame sequence. The spectral flattening can be a preprocessing method to compensate for high-frequency attenuation of the acoustic signal during transmission. The framing / windowing can be a method of dividing the long-time acoustic signal into short-time signal segments and applying a window function (Hamming window) to these short-time signal segments. The resulting acoustic frame sequence can be a sequence of short-time signal segments arranged in chronological order.

[0040] The second step is to perform the following steps for each acoustic frame in the above acoustic frame sequence: The third step involves performing a spectral transformation on the aforementioned sound wave frame to obtain a nonlinear sensing frequency scale. This nonlinear sensing frequency scale can simulate the characteristics of human hearing, mapping linear frequencies to a nonlinear sensing space.

[0041] As an example, firstly, a Fast Fourier Transform can be performed on each acoustic frame to obtain a linear spectrum. Then, the linear frequencies are mapped to a Mel frequency scale using a Mel scale formula to obtain a nonlinear sensing frequency scale. Furthermore, a pre-set triangular bandpass filter bank is used. For example, 26 Mel filters can be set to cover the entire analysis frequency band.

[0042] The fourth step is to extract the cepstral coefficients of the aforementioned nonlinear sensing frequency scale to obtain the Mel-frequency cepstral coefficients. These cepstral coefficients can be obtained by taking the logarithm of the spectrum and then performing a discrete cosine transform.

[0043] The fifth step is to perform differential processing on the above Mel-Cepstral coefficients to obtain the first-order difference and the second-order difference.

[0044] Step 6: Based on the aforementioned Mel-Cepstral Coefficients, first-order differences, and second-order differences, generate the temporal characteristics of the aforementioned acoustic frame. These temporal characteristics can be feature vectors used to characterize the spectral envelope and dynamic properties of the acoustic signal. In practice, the Mel-Cepstral Coefficients, first-order differences, and second-order differences can be concatenated as vectors to obtain the temporal characteristics.

[0045] Step 7: Based on the aforementioned nonlinear sensing frequency scale, construct a two-dimensional distribution map of each sound wave frame in the time and frequency dimensions. This two-dimensional distribution map can be an image of energy distribution in the time and frequency dimensions.

[0046] As an example, firstly, a Mel frequency filter bank is constructed based on a nonlinear sensing frequency scale. Secondly, for each acoustic frame, the spectrum is passed through the Mel frequency filter bank to calculate the energy value of each filter output channel, obtaining the energy distribution vector of the acoustic frame. Finally, the obtained energy distribution vectors are arranged in chronological order to obtain a two-dimensional matrix, which serves as a 2D distribution map.

[0047] Step 8: Perform range compression and normalization on the above two-dimensional distribution map to generate the two-dimensional image features of the above acoustic signal. The range compression can be a method of dynamically compressing the image pixel values ​​(e.g., logarithmic compression).

[0048] The ninth step is to integrate the above two-dimensional image features and temporal feature sequences to obtain acoustic wave feature information.

[0049] Step 103: Based on the dual-branch neural network model, feature extraction is performed on the vibration feature information and sound wave feature information to obtain the vibration feature vector and sound wave feature vector.

[0050] In some embodiments, the execution entity may perform feature extraction on the vibration feature information and the acoustic feature information based on a dual-branch neural network model to obtain vibration feature vectors and acoustic feature vectors.

[0051] In some optional implementations of certain embodiments, the execution entity may perform feature extraction on the vibration feature information and the acoustic feature information based on a dual-branch neural network model to obtain vibration feature vectors and acoustic feature vectors, which may include the following steps: The first step involves inputting the temporal feature sequence from the aforementioned vibration feature information into the first local feature extraction network of the aforementioned dual-branch neural network model, outputting a vibration local feature sequence. This vibration local feature sequence can be a feature sequence that preserves the local temporal correlation of the vibration signal. The first local feature extraction network can be a neural network structure (one-dimensional convolutional neural network) used to extract local temporal patterns of the vibration signal. In practice, the first local feature extraction network can include two convolutional blocks, each block comprising a one-dimensional convolutional layer (kernel size = 3, stride = 1, padding = 1), a batch normalization layer, a ReLU activation function, and a max-pooling layer (pooling size = 2). The first convolutional block has 32 output channels, and the second convolutional block has 64 output channels.

[0052] As an example, firstly, the temporal feature sequence in the feature information can be reconstructed into a two-dimensional matrix form. Secondly, the reconstructed feature information is input into the first local feature extraction network and processed by convolution to obtain the vibration local feature sequence.

[0053] The second step involves inputting the aforementioned local vibration feature sequence into a bidirectional temporal dependency network for the vibration branch, outputting a vibration feature vector. This bidirectional temporal dependency network can be a neural network structure (Long Short-Term Memory network) used to capture long-term forward and backward dependencies of the signal. The vibration feature vector can be the feature vector output by the vibration branch. In practice, the bidirectional temporal dependency network comprises two layers of bidirectional LSTM, with the hidden layer dimension set to 128 in each direction. The hidden states in the final step of the bidirectional LSTM (the concatenation of the final states in the forward and backward directions) form a 256-dimensional vector.

[0054] The third step is to determine the eigenvectors of the Mel cepstral coefficients and the eigenvectors of the Mel spectrum in the above acoustic wave characteristic information.

[0055] The fourth step involves inputting the aforementioned Mel-Cepstral Coefficient feature vectors into the second local feature extraction network of the acoustic branch, outputting Mel-Cepstral feature vectors. This second local feature extraction network can be a neural network structure (one-dimensional convolutional neural network) used to extract local temporal patterns of Mel-Cepstral Coefficient features. In practice, this second local feature extraction network may include two convolutional blocks and a global average pooling layer. Each block includes a one-dimensional convolutional layer (kernel size = 3, stride = 1, padding = 1), a batch normalization layer, a ReLU activation function, and a max-pooling layer (pooling size = 2). The first convolutional block has 64 output channels, and the second convolutional block has 128 output channels.

[0056] The fifth step involves inputting the aforementioned Mel spectrogram feature vector into a spatial feature extraction network to obtain the spectrogram feature vector. This spatial feature extraction network can be a neural network structure used to extract spatial structural features from a two-dimensional image. In practice, this network may include three convolutional blocks, each containing two 3×3 convolutional layers with 32, 64, and 128 channels respectively, followed by a 2×2 max-pooling layer. Finally, a fully connected layer is added, resulting in an output dimension of 256.

[0057] The sixth step is to linearly map the above spectral feature vectors into query vectors, and to map the above Mel-Cepstral feature vectors into key vectors and value vectors.

[0058] Step 7: Determine the similarity information between the key vector and the query vector, and normalize this similarity information to obtain the attention weights. The similarity information can be a measure of the degree of matching between the query vector and the key vector. In practice, similarity calculation can be... The above. It can be a 1×d dimensional vector (d is the feature dimension), i.e., the query vector. (The above...) It can be a 1×d dimensional vector, i.e., a key vector. Calculate their dot product (inner product) to obtain a scalar (a single numerical value). This scalar can be the similarity between the query vector and the key vector (the higher the similarity, the larger the value). The attention weight calculation can be... The above. This can be done through normalization. (The above...) This could be a scaling process. (The above...) It can be a feature dimension.

[0059] Step 8: Based on the aforementioned attention weights, perform a weighted summation on the above value vectors to obtain the Mel-Cepstral Weighted Feature Vector. This Mel-Cepstral Weighted Feature Vector can be a vector of Mel-Cepstral coefficient features after attention filtering.

[0060] The ninth step is to concatenate and fuse the above Mel cepstral weighted feature vector with the above spectral feature vector to generate the acoustic feature vector.

[0061] In some optional implementations of some embodiments, the above-described dual-branch neural network model is a dual-branch neural network model with a vibration branch and a sound wave branch; and the steps for training the above-described dual-branch neural network model include: The first step involves constructing the vibration branch and acoustic branch of the aforementioned dual-branch neural network model, resulting in an initial dual-branch neural network model. The vibration branch includes convolutional layers and a bidirectional temporal dependency network layer, while the acoustic branch includes a Mel-Cepstral Coefficients (MCC) network layer, a Mel-Spectrum network layer, and a cross-modal attention fusion layer. The initial dual-branch neural network model includes a modal attention layer and a classifier layer. The convolutional layers can be one-dimensional convolutional neural network layers. These layers can be used to extract local temporal features of the vibration signal. The bidirectional temporal dependency network layer can be a bidirectional long short-term memory (LSTM) network layer. This layer can be used to capture the forward and backward long-term dependencies of the vibration signal. The MCC network layer can be a one-dimensional convolutional neural network combined with global average pooling. The Mel-Spectrum network layer can be a two-dimensional convolutional neural network. This layer can be used to extract spatial structural features from the Mel-Spectrum. The cross-modal attention fusion layer can be an attention mechanism layer. The aforementioned modal attention layer can be a fully connected network layer that dynamically weights and fuses vibration feature vectors and acoustic feature vectors. The aforementioned classifier layer can be a Softmax classification layer used to output classification results of external damage behavior.

[0062] As an example, the constructed vibration branch comprises two one-dimensional convolutional blocks (each block including a convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer) and two bidirectional LSTM layers. The acoustic branch comprises a Mel-Cepstral Coefficients subnetwork (two one-dimensional convolutional blocks and global average pooling), a Mel-Spectrum subnetwork (three two-dimensional convolutional blocks and a fully connected layer), and a cross-modal attention fusion layer. The overall model also includes a modal attention layer (a small fully connected network) and a Softmax classification layer.

[0063] The second step involves acquiring a set of vibration signal samples and a set of acoustic signal samples for training. The vibration signal sample set can be a collection of vibration feature data including various external damage behaviors (e.g., mechanical excavation, manual hammering, and vehicle crushing). The acoustic signal sample set can be a collection of acoustic feature data acquired synchronously with the vibration signal samples.

[0064] Third, based on the above vibration signal sample set, perform the following vibration branch training steps: The first sub-step involves inputting at least one vibration signal sample from the aforementioned vibration signal sample set into the convolutional layer of the initial dual-branch neural network model to obtain a vibration signal feature vector corresponding to at least one vibration signal sample.

[0065] As an example, the vibration sequence (of shape) of multiple vibration samples in a batch (B is the batch size, T is the time step, and D is the feature dimension) Input to a one-dimensional convolutional layer, the output shape can be... The local features are used as the feature vector of the vibration signal.

[0066] The second sub-step involves inputting the vibration signal feature vector corresponding to at least one vibration signal sample into the bidirectional temporal dependency network layer of the initial dual-branch neural network model to obtain the vibration feature vector corresponding to at least one vibration signal sample. In practice, the local vibration feature sequence is expanded step-by-step and input into a two-layer bidirectional LSTM network. The output shape is... The vibration characteristic vector.

[0067] Fourth, based on the above acoustic signal sample set, perform the following acoustic branch training steps: The first sub-step involves inputting at least one acoustic signal sample from the aforementioned acoustic signal sample set into the Mel-Cepstral Coefficient Network Layer of the aforementioned initial dual-branch neural network model to obtain the Mel-Cepstral Coefficient Feature Vector corresponding to at least one acoustic signal sample, and to determine the Mel-Spectrum corresponding to the aforementioned at least one acoustic signal sample.

[0068] As an example, the Mel-frequency cepstral coefficient feature sequence (of shape ) of multiple acoustic samples in a batch is used. The input is fed into a one-dimensional convolutional layer of a Mel-frequency cepstral coefficient subnetwork. After two convolutional blocks and global average pooling, the output shape is... The eigenvectors of the Mel-Cepstral Coefficients. Simultaneously, the corresponding Mel-spectral data (shape: ...) are directly obtained from the acoustic wave samples. ).

[0069] The second sub-step involves inputting the Mel spectrogram corresponding to at least one acoustic signal sample into the Mel spectrogram network layer of the initial dual-branch neural network model to obtain the Mel spectrogram feature vector corresponding to at least one acoustic signal sample.

[0070] As an example, when a Mel spectrum is input into a two-dimensional convolutional neural network, after passing through three convolutional blocks and a fully connected layer, the output shape is as follows: The eigenvectors of the Mel spectrum.

[0071] The third sub-step involves inputting the Mel cepstral coefficient feature vector and the Mel spectrogram feature vector corresponding to at least one acoustic signal sample into the cross-modal attention fusion layer of the initial dual-branch neural network model to obtain the acoustic feature vector corresponding to at least one acoustic signal sample.

[0072] Fifth, based on the vibration feature vector corresponding to at least one vibration signal sample and the acoustic feature vector corresponding to at least one acoustic signal sample, determine the target loss function. This target loss function can be a loss function used to measure the difference between the model's prediction result and the true label.

[0073] The sixth step is to determine whether the initial dual-branch neural network model has achieved the preset optimization objective based on the aforementioned target loss function. The preset optimization objective may be loss function convergence, validation set accuracy reaching a threshold, or reaching a preset number of training epochs.

[0074] Step 7: In response to the initial bibranch neural network model achieving the preset optimization objective, the initial bibranch neural network model is adopted as the bibranch neural network model.

[0075] Step 8: In response to the initial dual-branch neural network model failing to achieve the preset optimization objective, the training parameters of the initial dual-branch neural network model are adjusted, and the adjusted initial dual-branch neural network model is used as the initial dual-branch neural network model to execute the model training steps again. These model training steps include: vibration branch training steps and sound wave branch training steps. The training parameters may include, but are not limited to, learning rate, batch size, and optimizer.

[0076] In some optional implementations of certain embodiments, the execution entity may determine the target loss function based on the vibration feature vector corresponding to the at least one vibration signal sample and the acoustic feature vector corresponding to the at least one acoustic signal sample, which may include the following steps: The first step is to concatenate the vibration feature vector corresponding to at least one vibration signal sample and the acoustic feature vector corresponding to at least one acoustic signal sample to obtain at least one acoustic-vibration fusion vector.

[0077] The second step is to input at least one of the above-mentioned acoustic-vibration fusion vectors into the modal attention layer of the above-mentioned initial dual-branch neural network model to obtain the vibration modal attention weights and acoustic modal attention weights.

[0078] The third step involves weighting and summing the vibration feature vector and the acoustic feature vector based on the aforementioned vibration mode attention weights and acoustic mode attention weights, to obtain at least one acoustic feature vector. The acoustic feature vector can be a fused feature vector after attention weighting.

[0079] As an example, the acoustic-vibration fusion vector can be input into a fully connected network (128 hidden layers, 2 output dimensions), and then the attention weights for the two modalities can be generated using the Softmax function. and ( Weighted fusion Among them, the above This could be the eigenvector corresponding to the vibration branch. (The above...) This could be the eigenvector corresponding to the acoustic wave branch. (The above...) This could be vibrational mode attention weights. (The above...) It can be acoustic modal attention weights. At least one acoustic feature vector is obtained.

[0080] Fourth, input at least one acoustic feature vector into the classifier layer of the initial dual-branch neural network model to obtain at least one cable external damage behavior information. The classifier layer can be a Softmax layer. This classifier layer can be used to output the probability of each cable external damage behavior type.

[0081] The fifth step involves comparing at least one of the aforementioned cable damage behaviors with the corresponding target cable damage behaviors to obtain the target loss function. In practice, this target loss function can be the cross-entropy loss function.

[0082] Step 104: Based on the attention mechanism, the vibration feature vector and the sound wave feature vector are weighted and fused to obtain the acoustic vibration feature vector.

[0083] In some embodiments, the execution entity may perform weighted fusion processing on the vibration feature vector and the acoustic feature vector based on an attention mechanism to obtain an acoustic vibration feature vector.

[0084] In some optional implementations of certain embodiments, the execution entity may perform weighted fusion processing on the vibration feature vector and the acoustic feature vector based on an attention mechanism to obtain the acoustic vibration feature vector, which may include the following steps: The first step is to concatenate the vibration feature vector and the acoustic feature vector to obtain a joint acoustic-vibration feature vector. This joint acoustic-vibration feature vector can be a combined feature vector possessing both vibration and acoustic features.

[0085] The second step is to input the aforementioned acoustic-vibration joint feature vector into the weight calculation network of the aforementioned dual-branch neural network model to obtain the first initial weight value and the second initial weight value, wherein the aforementioned first initial weight value is the initial weight value of the vibration mode and the aforementioned second initial weight value is the initial weight value of the acoustic mode.

[0086] As an example, the joint acoustic-vibration feature vector can be input into a fully connected network (128 hidden layers, 2 output dimensions, using the ReLU activation function) to obtain the first and second initial weight values.

[0087] The third step is to normalize the first initial weight value and the second initial weight value to obtain the vibration mode attention weight and the acoustic mode attention weight.

[0088] As an example, the first initial weight value and the aforementioned second initial weight value are normalized using the Softmax function to generate attention weights for the two modalities. and ( Weighted fusion Among them, the above This could be the eigenvector corresponding to the vibration branch. (The above...) This could be the eigenvector corresponding to the acoustic wave branch. (The above...) This could be vibrational mode attention weights. (The above...) It could be acoustic modal attention weights.

[0089] Fourth, based on the aforementioned vibration mode attention weights and acoustic mode attention weights, the aforementioned vibration feature vectors and acoustic feature vectors are multiplied element-wise to obtain weighted vibration feature vectors and weighted acoustic feature vectors. The aforementioned weighted vibration feature vector can be the feature vector obtained by weighting the vibration feature vectors with attention weights. The aforementioned weighted acoustic feature vector can be the feature vector obtained by weighting the acoustic feature vectors with attention weights.

[0090] The fifth step is to perform element-wise addition and fusion of the above-mentioned weighted vibration feature vector and the above-mentioned weighted acoustic feature vector to generate an acoustic vibration feature vector.

[0091] Step 105: Based on the acoustic vibration feature vector, output the cable external damage behavior identification result through the classifier.

[0092] In some embodiments, the aforementioned execution entity can output cable external damage behavior identification results based on the aforementioned acoustic vibration feature vectors through a classifier. The classifier can be a neural network structure used to map feature vectors to predefined behavior categories. The cable external damage behavior identification results can be confidence information of the external damage event type output by the model.

[0093] As an example, firstly, the acoustic vibration feature vector can be input into a classifier to obtain confidence information about the external damage behavior of the cable. In practice, the above classifier can adopt a two-layer fully connected network structure. The first fully connected layer can map the acoustic vibration feature vector to a 128-dimensional hidden layer, using the ReLU activation function to introduce a nonlinear transformation. To suppress overfitting, a Dropout layer can be added after this layer. The second fully connected layer can further map the 128-dimensional hidden features to... The output vector of dimension, where the above This can be the preset total number of external breach behavior categories. Finally, for The output vector is normalized using the Softmax function to obtain the probability distribution for each category, which serves as the result of cable external damage identification. The category corresponding to the maximum value in the probability distribution can be the type of external damage behavior determined by the model. This maximum value can also represent the model's confidence in that determination. The categories of external damage behaviors can be defined according to the actual application scenario and may include, but are not limited to: no event, mechanical excavation, hydraulic breaking, manual hammering, and vehicle crushing.

[0094] Step 106: Based on the acoustic and vibration signals, determine the location information corresponding to the cable external damage identification result, and send the cable external damage identification result and location information to the alarm platform to trigger an alarm.

[0095] In some embodiments, the executing entity can determine the location information corresponding to the cable external damage identification result based on the acoustic signal and the vibration signal, and send the cable external damage identification result and the location information to an alarm platform to trigger an alarm. The location information can be the spatial coordinates (distance) along the cable path from the point of external damage. The alarm platform can be an IoT cloud platform for receiving, processing, and pushing alarm information.

[0096] As an example, the first step is to determine the timestamp information of the acoustic and vibration signals to obtain the initial distance from the fiber optic sensing system. This initial distance is affected by the system's sampling interval and the accuracy of the light speed conversion; the positioning accuracy error is within [a certain range]. rice.

[0097] The second step introduces the Time Difference of Arrival (TDOA) algorithm. This TDOA algorithm is used in scenarios where multiple sensor monitoring points are deployed in critical areas (e.g., cable joints and road crossings). In key protected areas, at least three fiber optic sensor slaves with synchronous acquisition capabilities (or multiple virtual sensor channels on the same fiber optic cable) are deployed at intervals of <1km along the cable path to ensure time synchronization accuracy <1ms between all nodes.

[0098] When an external failure event occurs, the M (M≥3) nearest sensing nodes will successively detect the signal of the event. The signal feature segments detected by each node (e.g., the initial edge or first peak of the vibration signal) are extracted. The Generalized Cross-Correlation (GCC-PHAT) algorithm can be used to calculate the signal characteristics between any two nodes. and Relative time difference of received signal PHAT weighting can effectively suppress noise and reverberation, highlighting time delay characteristics. The calculation formula can be... Among them, the above and This could be the Fourier transform of the signal. (The above...) It can be a node Fourier transform of the received signal. (The above) It can be a node Fourier transform of the received signal. (The above) It can be a node The complex conjugate of the Fourier transform of the received signal. (The above...) and It can be a signal and The cross-power spectrum. The above. This could represent the inverse Fourier transform. (The above...) It can be for the purpose of making To obtain the maximum value The above. It could be a signal reaching a node. and nodes The time difference. The above. It can be a time delay variable.

[0099] To establish a geometric model for position calculation, we can assume that the coordinates of the sensing nodes (event sources) can be... , No. The known coordinates of a sensor node can be The propagation speed of sound / vibration waves in soil / cable medium is (On-site calibration is required; typical values ​​are approximately 300-1500 m / s, depending on soil type). For nodes and nodes We have the following equation: The above equations can be a hyperbola, with the event source located on a hyperbola with two nodes as foci. The overdetermined system of equations is solved using the least squares method to obtain the coordinates of the event source. The optimal estimate is obtained. By fusing information from more than three nodes, the measurement error of a single node can be reduced, improving the positioning accuracy to <5m.

[0100] The third step is to calculate the confidence score of the positioning result (ranging from 0 to 1) based on the number of nodes involved in the positioning, the signal-to-noise ratio, and the degree of dispersion of the intersection of each hyperbola.

[0101] The fourth step, after obtaining the location information, involves encapsulating the event information into a standard format data packet on an edge computing device. This alarm data packet is then published to the alarm platform (IoT cloud platform) via the MQTT protocol. Upon receiving the alarm information, the alarm platform performs data entry, real-time push notifications, multi-channel notifications, and work order generation. Maintenance personnel can then view the precise alarm location icon on a map within the interface (e.g., PC and mobile devices) for handling external damage events.

[0102] Based on the same inventive concept, embodiments of this application provide a method system for locating external damage to fiber optic dual-mode acoustic-vibration cables, comprising: The acquisition module is used to acquire sound wave signals and vibration signals; The memory is used to store the program for a method of locating external damage to a fiber optic acoustic-vibration dual-mode cable; The processor and memory can load and execute the program to implement a method for locating external damage to a fiber optic acoustic vibration dual-mode cable.

[0103] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0104] This application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as a method for locating external damage to a fiber optic acoustic-vibration dual-mode cable.

[0105] Computer storage media include, for example, USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media that can store program code.

[0106] Based on the same inventive concept, this application provides a smart terminal, including a memory and a processor. The memory stores a computer program that can be loaded and executed by the processor to locate external damage to a fiber optic acoustic vibration dual-mode cable.

[0107] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0108] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for locating external damage to a fiber optic dual-mode acoustic-vibration cable, characterized in that, include: Based on a single sensing fiber and a timing controller, acoustic and vibration signals along the cable are acquired, and the acoustic and vibration signals are spatiotemporally registered to obtain dual-mode signals. The dual-mode signal is subjected to feature processing to obtain vibration feature information and acoustic wave feature information. The steps of performing feature processing on the dual-mode signal to obtain vibration feature information and acoustic wave feature information include: The vibration signal in the dual-mode signal is decomposed into vibration signal components by multi-scale decomposition. The vibration signal components are denoised to obtain a denoised vibration signal. The denoised vibration signal is subjected to time-domain analysis to obtain time-domain feature information, wherein the time-domain feature information includes: root mean square, peak factor and zero-crossing rate; The denoised vibration signal is subjected to frequency domain transformation processing to obtain frequency domain feature information, wherein the frequency domain feature information includes: spectral centroid, spectral width, and subband energy ratio; The vibration characteristic information is obtained by integrating the time-domain feature information and the frequency-domain feature information; The acoustic wave signal in the dual-mode signal is subjected to acoustic wave feature processing to obtain acoustic wave feature information; Based on a dual-branch neural network model, feature extraction is performed on the vibration feature information and the sound wave feature information to obtain vibration feature vectors and sound wave feature vectors. The steps of extracting features from the vibration feature information and the sound wave feature information to obtain vibration feature vectors and sound wave feature vectors based on the dual-branch neural network model include: The temporal feature sequence in the vibration feature information is input into the first local feature extraction network of the dual-branch neural network model, and the vibration local feature sequence is output. The vibration local feature sequence is input into a bidirectional temporal dependency network of the vibration branch, and the vibration feature vector is output. Determine the Mel cepstral coefficient eigenvector and Mel spectrogram eigenvector in the acoustic wave feature information; The Mel-Cepstral Coefficient Feature Vector is input into the second local feature extraction network of the acoustic branch, and the Mel-Cepstral Feature Vector is output. The Mel spectrum feature vector is input into a spatial feature extraction network to obtain the spectrum feature vector; The spectral feature vector is linearly mapped to a query vector, and the Mel-Cepstral feature vector is mapped to a key vector and a value vector; The similarity information between the key vector and the query vector is determined, and the similarity information is normalized to obtain the attention weight; Based on the attention weights, the value vector is weighted and summed to obtain the Mel-Cepstral Weighted Feature Vector; The Mel cepstral weighted feature vector and the spectral feature vector are concatenated and fused to generate the acoustic feature vector; Based on the attention mechanism, the vibration feature vector and the sound wave feature vector are weighted and fused to obtain the acoustic vibration feature vector. Based on the acoustic and vibration feature vectors, a classifier outputs the identification results of cable external damage behavior. Based on the acoustic signal and the vibration signal, the location information corresponding to the cable external damage identification result is determined, and the cable external damage identification result and the location information are sent to the alarm platform to trigger an alarm.

2. The method for locating external damage to a fiber optic dual-mode acoustic-vibration cable according to claim 1, characterized in that, The steps of acquiring acoustic and vibration signals along a cable based on a single sensing fiber and a timing controller, and performing spatiotemporal registration processing on the acoustic and vibration signals to obtain dual-mode signals include: A first probe light and a second probe light are coupled into a sensing optical fiber using wavelength division multiplexing technology. The first probe light is used to excite and collect the signal of the sound wave, and the second probe light is used to excite and collect the signal of the vibration. The optical signal along the cable is synchronously sampled using a dual-channel synchronous data acquisition card and a single sensing optical fiber to obtain acoustic and vibration signals. Based on the synchronous trigger clock generated by the timing controller, the hardware timestamp information corresponding to the acoustic signal and the vibration signal is determined to obtain the acoustic signal data packet and the vibration signal data packet. In response to the dual-channel synchronous data acquisition card being in signal acquisition mode, hardware timestamp information is extracted from acoustic signal data packets and vibration signal data packets, and time-domain translation compensation is performed on the sampling sequence of the acoustic channel based on a fixed transmission delay difference to obtain an acoustic-vibration aligned signal, wherein the fixed transmission delay difference is the transmission delay difference between the acoustic signal and the vibration signal. For each acoustic vibration alignment signal point in the acoustic vibration alignment signal, a spatiotemporal tag is generated based on the hardware timestamp information and the spatial positioning information of the fiber optic sensing system, resulting in a spatiotemporal tag sequence. The spatiotemporal tag sequence and the acoustic-vibration alignment signal are encapsulated to obtain a dual-mode signal.

3. The method for locating external damage to a fiber optic dual-mode acoustic-vibration cable according to claim 1, characterized in that, The step of performing acoustic feature processing on the acoustic signal in the dual-modal signal to obtain acoustic feature information includes: The acoustic signal is subjected to spectral flattening and frame windowing to obtain an acoustic frame sequence; For each acoustic frame in the acoustic frame sequence, perform the following steps: The acoustic frame is subjected to spectral transformation to obtain a nonlinear sensing frequency scale; Extract the cepstral coefficients of the nonlinear sensing frequency scale to obtain the Mel-frequency cepstral coefficients; The Mel-Cepstral coefficients are differentially processed to obtain first-order and second-order differences; The temporal characteristics of the acoustic frame are generated based on the Mel-frequency cepstral coefficients, the first-order difference, and the second-order difference. Based on the nonlinear sensing frequency scale, a two-dimensional distribution map of each sound wave frame in the time and frequency dimensions is constructed. The two-dimensional distribution map is subjected to range compression and normalization to generate two-dimensional image features of the acoustic signal; The two-dimensional image features and the temporal feature sequence are integrated to obtain acoustic wave feature information.

4. The method for locating external damage to a fiber optic dual-mode acoustic-vibration cable according to claim 1, characterized in that, The bi-branch neural network model is a bi-branch neural network model with a vibration branch and a sound wave branch; The steps for training the dual-branch neural network model include: The vibration branch and acoustic branch of the dual-branch neural network model are constructed to obtain the initial dual-branch neural network model. The vibration branch includes a convolutional layer and a bidirectional temporal dependency network layer. The acoustic branch includes a Mel-Cepstral Coefficients Network Layer, a Mel-Spectrogram Network Layer, and a cross-modal attention fusion layer. The initial dual-branch neural network model includes a modal attention layer and a classifier layer. Obtain a set of vibration signal samples and a set of acoustic signal samples for training; Based on the vibration signal sample set, perform the following vibration branch training steps: At least one vibration signal sample from the vibration signal sample set is input into the convolutional layer of the initial dual-branch neural network model to obtain a vibration signal feature vector corresponding to at least one vibration signal sample. The vibration signal feature vector corresponding to the at least one vibration signal sample is input into the bidirectional temporal dependent network layer of the initial dual-branch neural network model to obtain the vibration feature vector corresponding to the at least one vibration signal sample. Based on the aforementioned acoustic signal sample set, the following acoustic branch training steps are performed: At least one acoustic signal sample from the acoustic signal sample set is input into the Mel-Cepstral Coefficient Network Layer of the initial dual-branch neural network model to obtain the Mel-Cepstral Coefficient Feature Vector corresponding to at least one acoustic signal sample, and to determine the Mel-Spectrum corresponding to the at least one acoustic signal sample. The Mel spectrogram corresponding to the at least one acoustic signal sample is input into the Mel spectrogram network layer of the initial dual-branch neural network model to obtain the Mel spectrogram feature vector corresponding to the at least one acoustic signal sample. The Mel cepstral coefficient feature vector and the Mel spectrogram feature vector corresponding to the at least one acoustic signal sample are input into the cross-modal attention fusion layer of the initial dual-branch neural network model to obtain the acoustic feature vector corresponding to at least one acoustic signal sample. Based on the vibration feature vector corresponding to the at least one vibration signal sample and the acoustic feature vector corresponding to the at least one acoustic signal sample, the target loss function is determined. Based on the target loss function, determine whether the initial dual-branch neural network model has achieved the preset optimization objective; In response to the initial dual-branch neural network model reaching the preset optimization target, the initial dual-branch neural network model is used as a dual-branch neural network model; In response to the initial bibranch neural network model failing to achieve the preset optimization target, the training parameters of the initial bibranch neural network model are adjusted, and the adjusted initial bibranch neural network model is used as the initial bibranch neural network model to perform the model training step again. The model training step includes a vibration branch training step and a sound wave branch training step.

5. The method for locating external damage to a fiber optic dual-mode acoustic-vibration cable according to claim 4, characterized in that, The step of determining the target loss function based on the vibration feature vector corresponding to the at least one vibration signal sample and the acoustic feature vector corresponding to the at least one acoustic signal sample includes: The vibration feature vector corresponding to the at least one vibration signal sample and the acoustic feature vector corresponding to the at least one acoustic signal sample are concatenated to obtain at least one acoustic-vibration fusion vector. The at least one acoustic-vibration fusion vector is input into the modal attention layer of the initial dual-branch neural network model to obtain vibration modal attention weights and acoustic modal attention weights; Based on the vibration mode attention weight and the acoustic mode attention weight, the vibration feature vector and the acoustic feature vector are weighted and summed to obtain at least one acoustic feature vector; The at least one acoustic feature vector is input into the classifier layer of the initial dual-branch neural network model to obtain at least one cable external damage behavior information; The at least one cable external damage behavior information is compared with the corresponding target cable external damage behavior to obtain the target loss function.

6. The method for locating external damage to a fiber optic dual-mode acoustic-vibration cable according to claim 1, characterized in that, The step of performing weighted fusion processing on the vibration feature vector and the sound wave feature vector based on the attention mechanism to obtain the acoustic vibration feature vector includes: The vibration feature vector and the sound wave feature vector are concatenated to obtain the joint sound-vibration feature vector. The joint acoustic-vibration feature vector is input into the weight calculation network of the dual-branch neural network model to obtain a first initial weight value and a second initial weight value, wherein the first initial weight value is the initial weight value of the vibration mode and the second initial weight value is the initial weight value of the acoustic mode. The first initial weight value and the second initial weight value are normalized to obtain the vibration mode attention weight and the acoustic mode attention weight; Based on the vibration mode attention weight and the acoustic mode attention weight, the vibration feature vector and the acoustic feature vector are multiplied element by element to obtain a weighted vibration feature vector and a weighted acoustic feature vector. The weighted vibration feature vector and the weighted acoustic feature vector are added and fused element by element to generate an acoustic vibration feature vector.

7. A fiber optic acoustic-vibration dual-mode cable external damage positioning system, characterized in that, include: The acquisition module is used to acquire sound wave signals and vibration signals; Memory, used to store computer programs; The processor and the program in the memory can be loaded and executed by the processor to implement the fiber optic acoustic vibration dual-mode cable external damage location method as described in any one of claims 1 to 6.

Images