A dual-branch acoustic-vibration fusion event recognition and positioning method based on DAS and AI

Abstract

The application discloses a kind of based on DAS and AI's double-branch sound vibration fusion event identification and positioning method, including synchronous acquisition sound and vibration data, construct sound identification branch and vibration positioning branch;Sound branch is extracted multi-scale joint feature by time-frequency domain feature fusion, the weight of dynamic adjustment CNN and BiLSTM is realized adaptive fusion, output event type and confidence;Vibration branch uses four-point space-time weighted optimization method to calculate initial coordinate, position fine-tuning is carried out in combination with ASTCN network, and the final event position is output.The results of fusion double-branch are used, and the probability distribution verification is carried out using dynamic likelihood ratio evaluation mechanism, and the final judgment result is output;The application realizes intelligent event classification by sound identification branch, high-precision event space-time positioning is realized by vibration positioning branch, and intelligent collaboration and system optimization of multi-modal data are realized through fusion and verification module, effectively break through the bottleneck of traditional technology in identification accuracy, positioning error and system robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of infrastructure monitoring, and particularly relates to a double-branch sound-vibration fusion event identification and positioning method based on DAS and AI. BACKGROUND

[0002] Distributed Acoustic Sensing (DAS) technology has been widely applied in the field of infrastructure monitoring in recent years, especially in scenarios such as cable tunnels, oil pipelines, railway lines, and border security. This technology uses optical fiber as a distributed sensor to detect phase changes in optical signals along the line, capturing sound and vibration signals in real time, thereby monitoring abnormal events such as cable failures, mechanical impacts, or intrusions. Compared with traditional point sensors, DAS has the characteristics of wide coverage (up to tens of kilometers), high sensitivity, and no need for additional power supply, and has become an important technical means in the field of intelligent monitoring.

[0003] However, despite the great potential of DAS technology in event monitoring, it still has many defects in practical application, which limits its further performance improvement. Specifically, the current status and defects of similar technology applications mainly reflect in the following aspects: insufficient event identification accuracy, especially in high-noise environments, it is difficult to effectively separate event signals from background noise, resulting in high false positive and false negative rates; poor positioning accuracy and stability, affected by signal attenuation along the fiber, multipath effect and environmental interference, especially in long-distance monitoring, the error is more obvious; insufficient environmental adaptability and robustness, poor adaptability to dynamic environmental changes such as temperature fluctuations, fiber aging and external electromagnetic interference, system performance declines over time; limited intelligence, relying on manual threshold setting or rules, lacking adaptive learning ability; low real-time and computational efficiency, existing deep learning models have high computational complexity, making it difficult to achieve real-time processing on edge devices, and the response speed is difficult to meet the rapid monitoring demand. The existence of these problems greatly restricts the application effect of DAS technology in complex scenarios.

[0004] Specifically, in the context of Distributed Acoustic Sensing (DAS) technology applied to infrastructure monitoring, existing technologies face a series of outstanding problems in event identification and positioning, as follows:

[0005] 1. Insufficient event recognition accuracy: Existing technologies have difficulty accurately distinguishing event signals from background noise in complex noise environments, such as traffic noise or wind noise in urban cable tunnels. Especially when multiple events overlap or transient events occur, the false positive rate and false negative rate increase significantly. Traditional signal processing methods have limited processing capacity for high-dimensional heterogeneous DAS data, making it difficult to capture the timing dynamic characteristics of sound and vibration. Traditional machine learning methods rely on manual feature extraction, which has poor adaptability, resulting in poor recognition performance.

[0006] 2. Poor positioning accuracy and stability: DAS systems usually calculate event locations based on signal arrival time differences or phase differences, but are affected by signal attenuation along the fiber line, multipath effects, and environmental interference, resulting in large positioning errors, especially in long-distance monitoring. Existing positioning algorithms are mostly based on static assumptions and lack dynamic tracking capabilities for spatiotemporal evolution of events, such as mobile intrusions or persistent faults, which can cause positioning results to drift or become ambiguous.

[0007] 3. Poor environmental adaptability and robustness: Existing technologies have poor adaptability to dynamic environmental changes, such as temperature fluctuations, fiber aging, or external electromagnetic interference, which can cause system performance to decline over time. In addition, relying on single-modal data processing (such as using only sound or vibration signals) is susceptible to interference in complex scenarios, such as strong winds that can mask sound signals and ground vibrations that can be confused with other mechanical activities, making it difficult to meet diverse monitoring needs.

[0008] 4. Limited intelligence: Current DAS systems rely on manually set thresholds or rules and lack adaptive learning capabilities, making it difficult to dynamically optimize model parameters based on real-time data. Although deep learning methods have been tried in DAS applications, they are limited by insufficient training data and single model structure, which significantly restricts their promotion in diverse scenarios.

[0009] Low real-time performance and computational efficiency: Existing deep learning models have high computational complexity, making it difficult to achieve real-time processing on edge devices, which limits the deployment capabilities of DAS systems in low-power scenarios. In addition, existing methods have slow response speed when processing large-scale DAS data, making it difficult to meet the actual needs of real-time monitoring and rapid response. SUMMARY

[0010] The technical problem to be solved by the present application is to provide a dual-branch sound-vibration fusion event recognition and positioning method based on DAS and AI to overcome the shortcomings of the existing technology.

[0011] The technical solution to solve the above technical problems is as follows: A dual-branch sound-vibration fusion event recognition and positioning method based on DAS and AI, comprising the following steps:

[0012] Step S01: collecting sound data from the DAS system; capturing multi-scale features including time domain features and frequency domain features from the sound data, weighting and adaptively fusing the time domain features and the frequency domain features to obtain joint features ;

[0013] Step S02: extracting high-level spatial features from the joint features , performing time sequence modeling on the feature sequence output by the CNN to generate time sequence features , dynamically adjusting the fusion weight of the CNN and the Bi-LSTM to obtain fused features , inputting the fused features into a fully connected layer, and outputting an event type and its confidence through a function, denoted as the output result of the sound recognition branch ;

[0014] Step S03: collecting vibration intensity data using the DAS system; pre-processing and standardizing the vibration intensity data to obtain standardized vibration intensity data ; calculating the initial positioning coordinates of the target point based on a four-point spatiotemporal weighted optimization positioning method ; fine-tuning the initial positioning coordinates to output the final event location , and obtaining the output result of the vibration positioning branch ;

[0015] Step S04: fusing and verifying the output result of the sound recognition branch and the output result of the vibration positioning branch to obtain a fusion probability distribution , performing consistency evaluation through a dynamic likelihood ratio to calculate a likelihood ratio , and generating a final fusion result by integrating the probability distributions of the two branches ;

[0016] Setting a judgment threshold , and comparing the likelihood ratio with the judgment threshold to determine the final .

[0017] ​The beneficial effects of the present application are: the present application aims to break through the bottleneck of traditional technology in complex environment event recognition accuracy, large space-time positioning error and poor system robustness. Through the "sound recognition branch", the intelligent event classification is realized, the "vibration positioning branch" realizes the high-precision event space-time positioning, and the "fusion and verification" module realizes the intelligent cooperation and system optimization of multi-modal data, thereby effectively breaking through the bottleneck of traditional technology in recognition accuracy, positioning error and system robustness, and providing an efficient, accurate and stable monitoring scheme for cable tunnels and other key infrastructures.

[0018] On the basis of the above technical scheme, the present application can also be improved as follows.

[0019] Further, step S01 specifically comprises the following steps:

[0020] Step S11: simultaneously extracting multi-scale features in time domain and frequency domain;

[0021] Step S111: extracting features of different time spans by using sliding windows or CNNs of different lengths; the output is represented as , wherein is the scale index.

[0022] Step S112: extracting features of different frequency resolutions by wavelet transform or STFT; the output is represented as .

[0023] Step S12: calculating attention weights in time domain and frequency domain respectively by a double-branch attention module, and weighting the features:

[0024] Step S121: calculating attention weights for the time domain features of each scale .

[0025] .

[0026] wherein, is a learnable weight matrix; the weighted feature is

[0027] .

[0028] wherein, represents element-wise multiplication.

[0029] Step S122: calculating attention weights for the frequency domain features of each scale .

[0030] .

[0031] Weighted feature:

[0032] ;

[0033] wherein, denotes element-wise multiplication;

[0034] Step S13: The scale weights are adaptively adjusted according to the input signal characteristics, the weighted features of multiple scales are fused, and time-frequency joint optimization is realized:

[0035] Step S131: The weighted features in the time domain and the frequency domain are respectively weighted and summed:

[0036] ;

[0037] ;

[0038] wherein, and are learnable scale weights, representing the contribution of each scale;

[0039] Step S132: The time domain and frequency domain features are integrated through a joint attention module to obtain joint features :

[0040] .

[0041] Further, step S02 specifically includes the following steps:

[0042] Step S21: The CNN extracts high-level spatial features from the joint features , which is realized through multi-layer convolution operation, and the calculation formula is as follows:

[0043] ;

[0044] ;

[0045] The parameters in the formula are as follows:

[0046] input features of the first layer (for the first layer, );

[0047] convolution kernel of the first layer, which is usually a four-dimensional tensor , wherein is the height and width of the convolution kernel, is the number of input channels, is the number of output channels;

[0048] convolution operation, which calculates the sliding window convolution of the input features and the convolution kernel;​​

[0049] The first layer bias term, shape is ;

[0050] Activation function;

[0051] The output feature map of the first layer, shape is , wherein Determined by the input size, convolution kernel size and stride;

[0052] Step S22: Bi-LSTM time series modeling of the feature sequence output by CNN, generating time series features ; The calculation process is as follows:

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] ;

[0058] ;

[0059] ;

[0060] wherein, and Respectively, the forward and backward hidden state; Indicates element-wise multiplication;

[0061] Step S23: According to the space-time characteristics of the input signal, dynamically adjust the fusion weight of CNN and Bi-LSTM, improve the adaptability of feature fusion; The calculation formula is as follows:

[0062] ;

[0063] wherein, adaptive weight Is calculated by the following formula:

[0064] ;

[0065] wherein, and Are learnable parameters;

[0066] Step S24: Extract features from the CNN middle layer and interact with the Bi-LSTM hidden state to obtain , the calculation formula is as follows:

[0067] ;

[0068] in, CNN Layer features, For Bi-LSTM in time The hidden state of

[0069] Step S25: or Input the fully connected layer of CNN, through The function outputs the event type and its confidence, which is recorded as the output result of the sound recognition branch .

[0070] Furthermore, step S131 further includes:

[0071] Use a lightweight neural network to predict the best scale combination based on the global characteristics of the input signal:

[0072] ;

[0073] in, are the original input features, is a global pooling operation, is the scale selection probability;

[0074] use Update scale weights and , to achieve dynamic optimization.

[0075] Furthermore, step S03 specifically includes the following steps:

[0076] Step S31: Preprocess and standardize the vibration intensity data to obtain standardized vibration intensity data ;

[0077] Step S32: Using the four-point spatiotemporal weighted optimization positioning method, the spatiotemporal information of four adjacent measurement points is used to calculate the position coordinates of the target point through weighted optimization. ;

[0078] Step S321: Obtain data from four adjacent test points, marked as , the position coordinates of each test point are represent;

[0079] Step S322: define signal parameters, the initial propagation speed of the signal; the time of signal arrival at the i-th test point; the time of signal arrival at the i-th test point;

[0080] Step S323: finding the earliest arrival time from the signal arrival times at the four test points, as follows:

[0081]

[0082] Step S324: calculating the weighting factor for each test point, as follows:

[0083]

[0084] wherein, is the signal strength or reliability factor of the i-th test point; is the attenuation coefficient; is the time of signal arrival at the i-th test point; is the earliest arrival time among the signal arrival times at the four test points; Step S325: using the weighting factor and the test point position, calculating the initial positioning coordinates of the target point, as follows:

[0085]

[0086]

[0087] wherein, is the position coordinate of the i-th test point; is the weighting factor; is the sum of the weighted positions of all test points; is the sum of all weights, for normalization; is the initial positioning coordinates of the target point;

[0088] Step S33: performing wavelet transform on the normalized signal, decomposing into multi-scale time-frequency features, as follows:

[0089]

[0090] wherein, is the wavelet base function, and are the scale and displacement parameters, respectively;

[0091] ​​​​​​​​​​​​​​​Step S34: Fine-tune the initial positioning coordinates by ASTCN The calculation formula is as follows:

[0092] ;

[0093] In the formula, is the initial position calculated by the four-point spatiotemporal weighted optimization method; is the final event position; is the calibration offset, calculated by the dynamic calibration function The formula is as follows:

[0094] ;

[0095] In the formula, is the feature vector generated by the multi-scale spatiotemporal correction; is the real-time monitored environmental noise level.

[0096] Further, the preprocessing and standardization steps in step S03 include the following specific steps:

[0097] Wavelet threshold denoising technology is used to process the original vibration signal ;

[0098] The vibration intensity is standardized, and the calculation formula is as follows:

[0099] ;

[0100] In the formula, is the mean value of the vibration intensity, is the standard deviation, is the standardized vibration intensity data.

[0101] Further, step S04 specifically includes the following steps:

[0102] Step S41: Map the event type probability of the sound recognition branch and the spatiotemporal features of the vibration positioning branch to a unified event hypothesis space to form a joint probability distribution, and the formula is as follows:

[0103] ;

[0104] In the formula, is the output result of the sound recognition branch; is the output result of the vibration positioning branch; is the event prior probability; is the normalization factor to ensure the validity of the probability distribution;

[0105] Step S42: Introduce a dynamic likelihood ratio to quantify the consistency of the sound and vibration branch outputs and identify potential conflicts, and the formula is as follows:

[0106] ;

[0107] where, is the assumption of consistent two-branch outputs; is the assumption of inconsistent two-branch outputs; is the likelihood ratio, the larger the value, the higher the consistency;

[0108] Step S43: Integrate the two-branch probability distribution to generate the final fusion result , the formula is:

[0109] ;

[0110] where, is the branch weight, dynamically adjusted according to data quality; is the normalization function, to ensure ;

[0111] Step S44: Dynamically adjust the decision threshold according to the likelihood ratio , the formula is as follows:

[0112] ;

[0113] where, is the basic threshold; is the adjustment coefficient; is the standard deviation of environmental noise;

[0114] Compare the likelihood ratio with the decision threshold :

[0115] When , the system accepts the current as the final decision result, without updating;

[0116] When , the system triggers the secondary analysis mechanism to re-evaluate the data and further update .

[0117] Further, step S41 also includes: adaptive Bayesian fusion network, dynamically optimizes the fusion process through environmental adaptive factor, the formula is as follows:

[0118] ;

[0119] where, is the environmental adaptive factor, defined as:

[0120] ;

[0121] where, is the environmental feature vector; are parameters optimized through machine learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0122] Figure 1 This is a diagram of the time-frequency joint modeling architecture of the present invention;

[0123] Figure 2 is an execution diagram of the DSTCO of the present invention;

[0124] Figure 3 This is a flow chart of the four-point spatiotemporal weighted optimization positioning method of the present invention;

[0125] Figure 4 This is a technical module diagram of the multi-scale spatiotemporal correction of the present invention;

[0126] Figure 5 Schematic diagram of the Adaptive Spatiotemporal Calibration Network (ASTCN) of the present invention;

[0127] Figure 6 This is a flowchart of the intelligent collaboration of the fusion and verification modules of the present invention. DETAILED DESCRIPTION

[0128] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0129] like Figures 1-6 As shown in Example 1, a dual-branch vibroacoustic fusion event recognition and positioning method based on DAS and AI includes the following steps:

[0130] Step S01: Collect sound data from the DAS system (such as cable discharge sound, mechanical impact sound, and human activity sound, stored in WAV format); sound signals have complex multi-scale characteristics in the time and frequency dimensions, such as short-term high-frequency transients (such as discharge sound) and long-term low-frequency continuous signals (such as mechanical vibration). Traditional attention mechanisms usually operate on a single scale or a single dimension (time domain or frequency domain), making it difficult to fully capture these characteristics, resulting in limited recognition performance in complex scenarios; capture multi-scale features including time domain features and frequency domain features from the sound data, weight and adaptively fuse the time domain features and frequency domain features to obtain joint features To address the multi-scale nature of sound signals, MSTFAA introduces a dynamic scale selection mechanism. Using a lightweight neural network, it predicts the optimal scale combination and adaptively adjusts the weights of time and frequency domain features. Compared to traditional fixed-scale methods, this mechanism dynamically optimizes feature fusion based on the input signal characteristics, improving the model's generalization and adaptability in diverse scenarios. This intelligent design demonstrates remarkable flexibility in handling complex sound events and demonstrates significant originality.

[0131] Step S02: Extract high-level spatial features from the joint features by a CNN (Convolutional Neural Network) module , perform temporal modeling on the feature sequence output by CNN through a Bi-LSTM (Bidirectional Long Short-Term Memory Network) module to generate temporal features , dynamically adjust the fusion weight of CNN and Bi-LSTM according to the spatio-temporal characteristics of the input signal to obtain the fusion features , input the fusion features into the fully connected layer, and output the event type and its confidence through function, denoted as the sound recognition branch output result ; Dynamic Spatio-Temporal Collaborative Optimization (DSTCO) is an adaptive feature fusion technology that dynamically adjusts the fusion weight of Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory Network (Bi-LSTM) output according to the spatio-temporal characteristics of the sound signal. Traditional methods usually use fixed fusion strategy, while DSTCO introduces adaptive weight , dynamically optimizes the feature fusion process according to the characteristics of the input signal, thereby significantly improving the adaptability and recognition accuracy of the model for different sound events.

[0132] Step S03: Collect vibration intensity data (stored in VIB format) using DAS system, and realize high-precision spatio-temporal positioning of event occurrence position through advanced signal processing and optimization algorithm. This branch effectively overcomes the challenges of optical fiber signal attenuation and environmental interference; pre-process and standardize the vibration intensity data to obtain standardized vibration intensity data ; adopt four-point spatio-temporal weighted optimization positioning method, use the spatio-temporal information (position and signal arrival time) of four adjacent measurement points to calculate the initial positioning coordinates of the target point through weighted optimization ; through the "distance-time double weighting" mechanism , consider the time delay and signal characteristics at the same time, which is more comprehensive than the traditional method based only on distance, and improves the robustness. The weight design skillfully quantifies the influence of time delay as nonlinear attenuation, giving priority to early signals and improving noise resistance. The introduction of makes the method dynamically adjustable according to signal quality, and the weighted average formula maintains low computational complexity, suitable for real-time applications. Use four test points instead of the traditional three to increase redundant information and effectively resist single measurement error; fine-tune the initial positioning coordinates through the adaptive spatio-temporal calibration network (ASTCN) to output the final event position , and obtain the vibration positioning branch output result vibration positioning branch output result including vibration intensity, position, time; break through the limitations of traditional fixed calibration method, realize dynamic self-adaptive calibration through the joint input of multi-scale features and environmental noise level. Introduce intelligent optimization mechanism, so that the system can still maintain high precision positioning in the scene of poor signal quality or strong interference.

[0133] To ensure the high consistency of the output results of the sound recognition branch and the vibration positioning branch, the application designs a fusion and verification module, which realizes the deep cooperation of multi-modal data through probability fusion, conditional probability calculation and weight adaptive adjustment.

[0134] Step S04: input the sound recognition branch output result and the vibration positioning branch output result into the Bayesian network fusion mechanism for fusion verification to obtain a fusion probability distribution , evaluate the consistency through dynamic likelihood ratio (Likelihood Ratio, LR), calculate the likelihood ratio , and generate the final fusion result ; The result of event recognition is a probability value, mainly reflecting the confidence of the event category, and does not directly contain the positioning result, which indirectly affects the calculation of ;

[0135] Set a judgment threshold , and compare the likelihood ratio with the judgment threshold :

[0136] When , the system accepts the current as the final judgment result without updating;

[0137] When , the system triggers the secondary analysis mechanism to re-evaluate the data and further update .

[0138] The application aims to break through the bottlenecks of traditional technology in complex environment, such as insufficient event recognition accuracy, large time and space positioning error and poor system robustness. Through the "sound recognition branch" to realize intelligent event classification, the "vibration positioning branch" to realize high-precision event time and space positioning, and the "fusion and verification" module to realize intelligent cooperation and system optimization of multi-modal data, the application effectively breaks through the bottlenecks of traditional technology in recognition accuracy, positioning error and system robustness, and provides an efficient, accurate and stable monitoring scheme for cable tunnels and other key infrastructures.

[0139] Embodiment 2, this embodiment is a further improvement made on the basis of Embodiment 1, which is as follows:

[0140] Step S01 specifically includes the following steps:

[0141] Step S11: Simultaneously extract multi-scale features in time domain and frequency domain, covering the diversification characteristics of sound signals. Simultaneously extract multi-scale features in time domain and frequency domain, covering the diversification characteristics of sound signals. In order to capture the multi-scale characteristics of sound signals, first extract features in time domain and frequency domain respectively:

[0142] Step S111: Multi-scale feature extraction:

[0143] Time domain multi-scale features: use different length sliding windows or CNN (multi-layer convolutional neural network) to extract features of different time span; for example, short window captures transient events, long window analyzes continuous signals; the output is represented as , where is the scale index;

[0144] Step S112: Frequency domain multi-scale features: extract features of different frequency resolution through wavelet transform or STFT (multi-resolution short-time Fourier transform); for example, low resolution captures low frequency trend, high resolution focuses on high frequency details; the output is represented as ; these features provide rich input for subsequent attention allocation;

[0145] Step S12: To realize the cooperative optimization of time domain and frequency domain, through a double-branch attention module, calculate the attention weight in time domain and frequency domain respectively, and weight the features:

[0146] Step S121: Time domain attention: calculate the attention weight for each scale of time domain features ;

[0147] ;

[0148] where, is a learnable weight matrix; the weighted feature is

[0149] ;

[0150] where, represents element-wise multiplication;

[0151] Step S122: Frequency domain attention: calculate the attention weight for each scale of frequency domain features ;

[0152] ;

[0153] Weighted features:

[0154] ;

[0155] where, denotes element-wise multiplication; in this way, the model can focus on key areas in the time and frequency domains.

[0156] MSTFAA is a unique time-frequency dual-branch attention module that calculates attention weights in the time and frequency domains, respectively, to achieve collaborative optimization of spatial and temporal information. Time-domain attention focuses on multi-scale dynamic changes, and frequency-domain attention focuses on multi-resolution frequency patterns. The dual-branch design enables the model to capture both transient events and high-frequency details, improving recognition accuracy and robustness in complex scenarios, and demonstrating the forward-looking and originality of the technology.

[0157] Step S13: adaptively adjust the scale weight according to the input signal characteristics to improve the generalization ability of the model; fuse the weighted features of multiple scales and realize time-frequency joint optimization:

[0158] Step S131: scale fusion: weighted sum of weighted features in time and frequency domains:

[0159] ;

[0160] ;

[0161] where, and are learnable scale weights, representing the contribution of each scale;

[0162] Step S132: time-frequency joint feature: integrate time and frequency domain features through a joint attention module to obtain a joint feature :

[0163] ;

[0164] Here, can be a simple weighted sum or a more complex multi-head self-attention mechanism.

[0165] The obtained using this mechanism can be used as the input feature of the first layer of the convolutional layer in the "CNN-BiLSTM sound event recognition fusion architecture". .

[0166] To address the multi-scale nature of acoustic signals, MSTFAA introduces a dynamic scale selection mechanism, predicting the optimal scale combination through a lightweight neural network and adaptively adjusting the time and frequency domain feature weights. Compared to traditional fixed scale methods, this mechanism dynamically optimizes feature fusion according to the characteristics of the input signal, enhancing the model's generalization ability and adaptability in diverse scenarios. This intelligent design exhibits outstanding flexibility when dealing with complex sound events, demonstrating significant originality.

[0167] Embodiment 3, this embodiment is a further improvement based on embodiment 2, which is as follows:

[0168] Step S02 specifically includes the following steps:

[0169] Step S21: CNN extracts high-level spatial features from the joint features through multi-layer convolution operations, with the calculation formula as follows:

[0170] ;

[0171] ;

[0172] The parameters in the formula are as follows:

[0173] : input features of the layer (for the first layer, );

[0174] : convolution kernel (weight matrix) of the layer, usually a four-dimensional tensor , where is the height and width of the convolution kernel, is the number of input channels, is the number of output channels;

[0175] convolution operation, calculating the sliding window convolution of input features and convolution kernel;

[0176] bias term of the layer, shape ;

[0177] activation function, here ReLU function, defined as , used to introduce nonlinearity;

[0178] output feature map of the layer, shape , where Determined by the input size, convolution kernel size and stride;

[0179] Step S22: Bi-LSTM performs time series modeling on the feature sequence output by CNN to generate time series features ; The calculation process is as follows:

[0180]

[0181]

[0182] in, and They are the forward and backward hidden states respectively, achieving bidirectional capture of temporal dependencies; Represents element-wise multiplication;

[0183] Step S23: Dynamic spatiotemporal collaborative optimization mechanism (DSTCO): Dynamically adjust the fusion weights of CNN and Bi-LSTM according to the spatiotemporal characteristics of the input signal to improve the adaptability of feature fusion; the calculation formula is as follows:

[0184]

[0185] in, and To learn parameters, ensure the intelligence and efficiency of the fusion process;

[0186] Figure 2 The implementation process of DSTCO is dynamic spatiotemporal collaborative optimization (DSTCO), which is an adaptive feature fusion technology that dynamically adjusts the fusion weights of the convolutional neural network (CNN) and the bidirectional long short-term memory network (Bi-LSTM) output according to the spatiotemporal characteristics of the sound signal. Traditional methods usually adopt a fixed fusion strategy, while DSTCO introduces adaptive weights. , dynamically optimizes the feature fusion process according to the characteristics of the input signal, thereby significantly improving the model's adaptability and recognition accuracy to different sound events.

[0187] Step S24: Extract features from the CNN middle layer and interact with the Bi-LSTM hidden state to enhance the integration effect of spatiotemporal information. , the calculation formula is as follows:

[0188]

[0189] in, CNN Layer features, For Bi-LSTM in time The hidden state of

[0190] Cross-layer feature interaction (CLFI) is a novel feature integration mechanism that directly interacts the spatial features of the intermediate layers of CNN with the temporal hidden states of Bi-LSTM to realize multi-level spatio-temporal information fusion. Traditional models usually perform feature fusion at a single level, while CLFI enhances the model's ability to capture complex spatio-temporal patterns in sound signals through cross-layer design;

[0191] Step S25: obtaining the feature vector of the input sound signal or The fully connected layer of the input CNN outputs the event type and its confidence through the function, denoted as the sound recognition branch output result .

[0192] Embodiment 4, which is a further improvement based on any one of embodiments 2 or 3, is as follows:

[0193] Step S131 further includes:

[0194] A scale selection module is used to allow the model to adapt to different input signals:

[0195] Scale selection prediction: a lightweight neural network is used to predict the best scale combination based on the global features of the input signal:

[0196]

[0197] wherein, is the original input feature, is the global pooling operation, is the scale selection probability;

[0198] Adaptive adjustment: the scale weight is updated to and to achieve dynamic optimization. This mechanism ensures that the model can select the most suitable feature scale in different scenarios.

[0199] Embodiment 5, which is a further improvement based on embodiment 1, is as follows:

[0200] The vibration positioning branch uses the vibration intensity data collected by the DAS system (stored in VIB format) to achieve high-precision spatio-temporal positioning of the event occurrence location through advanced signal processing and optimization algorithms. This branch effectively overcomes the challenges of optical fiber signal attenuation and environmental interference;

[0201] Step S03 specifically includes the following steps:

[0202] Step S31: pre-processing and standardizing the vibration intensity data to obtain standardized vibration intensity data ;

[0203] The present technology proposes an innovative "four-point spatiotemporal weighted optimization positioning method", aiming to utilize the spatiotemporal information (position and signal arrival time) of four adjacent measurement points to calculate the position coordinates of the target point through weighted optimization. The core lies in introducing a composite weight to comprehensively consider the influence of signal arrival time delay, thereby optimizing the inaccurate problems caused by measurement errors in traditional positioning methods. By weighted calculation of the vibration intensity and time delay of the four adjacent measurement points, high-precision positioning of the event position is realized; see Figure 3 the flowchart of the four-point spatiotemporal weighted optimization positioning method;

[0204] Step S32: Adopting the four-point spatiotemporal weighted optimization positioning method, the position coordinates of the target point are calculated through weighted optimization;

[0205] Step S321: Obtain data from four adjacent test points, respectively marked as , the position coordinates of each test point are represented by ;

[0206] Step S322: Define signal parameters, as the initial propagation speed of the signal; as the time of signal arrival at the th test point; this step collects the original data required for positioning, and the test point positions are used for subsequent calculation of target coordinates, and are used for analysis of the spatiotemporal characteristics of signal propagation;

[0207] Step S323: Find the earliest arrival time from the signal arrival times of the four test points, the formula is:

[0208]

[0209] is the minimum value among all , serving as the reference for time delay calculation. It reflects the test point where the signal arrives the fastest, usually implying that this point has the least measurement disturbance and higher reliability.

[0210] Step S324: Calculate the weighting factor for each test point , the calculation formula is as follows:

[0211]

[0212] In the formula, is the Signal strength or reliability factor at each test point; is the attenuation coefficient, a constant used to control the decay speed of the weight due to time delay; The signal reaches The time of each test point; is the earliest arrival time of the signals at the four test points;

[0213] Indicates the The delay of each test point relative to the earliest arrival time, the greater the delay, The smaller, The lower the weight. It is an exponential decay function, ensuring that the test points with earlier signal arrival times receive higher weights; Introducing additional flexibility, allowing weights to be adjusted based on signal quality (e.g., strength). Dynamically reflects the measurement reliability of each test point, with points with smaller time delays contributing more;

[0214] Step S325: Using weighting factors and test point locations , calculate the initial positioning coordinates of the target point , the calculation formula is as follows:

[0215]

[0216] Where, For the The position coordinates of the test points; is the weighting factor; is the sum of the weighted positions of all test points; is the sum of all weights, used for normalization; is the initial positioning coordinate of the target point;

[0217] This formula is a mathematical expression of weighted average, ensuring that test points with high weights contribute more to the final position; the normalization factor Make the results independent of the absolute value of the weight and only reflect the relative contribution;

[0218] By weighted averaging and integrating the information of the four test points, the optimal estimated position of the target point is obtained.

[0219] Through the "distance-time double weighting" mechanism , taking into account the time delay and signal characteristics , which is more comprehensive and has improved robustness compared to traditional distance-based methods. The weight design of the formula cleverly quantifies the influence of time delay as nonlinear attenuation, prioritizes early signals, and improves noise resistance. The introduction of the formula makes the method dynamically adjustable according to signal quality, and the weighted average formula maintains low computational complexity, suitable for real-time applications. Four test points are used instead of the traditional three, increasing redundant information and effectively resisting single measurement errors.

[0220] To further optimize the positioning results, the technology introduces a "multi-scale space-time correction" mechanism to improve signal quality through time-frequency analysis and interference suppression.

[0221] Step S33: Wavelet transform time-frequency analysis: wavelet transform is performed on the normalized signal to decompose it into multi-scale time-frequency features, and the calculation formula is as follows:

[0222]

[0223] where, is the wavelet basis function, and are the scale and displacement parameters, respectively. Through this step, the system can capture comprehensive features of the vibration signal from low-frequency trends to high-frequency transients.

[0224] This is a two-dimensional function that reflects the energy distribution of the signal in the "scale-time" domain. It can effectively reflect the dynamic change characteristics of the signal at multiple scales. Based on this function, the main energy position, frequency band energy distribution, and statistical features can be extracted to construct a multi-scale feature vector, which is used as the key input of ASTCN, effectively realizing space-time feature correction and positioning accuracy improvement.

[0225] Beamforming technology is based on directional enhancement algorithm to suppress multipath effect and environmental noise interference, further optimize feature representation, and ensure the stability of positioning results.

[0226] Referring to Figure 4 , the technical module diagram of multi-scale space-time correction is shown;

[0227] This scheme breaks through the limitations of traditional Fourier transform in time-frequency locality, uses wavelet transform to realize multi-resolution analysis, and more accurately captures the dynamic changes of the vibration signal. Combined with beamforming technology, it effectively suppresses interference and improves the robustness of features, significantly better than traditional single-scale analysis methods.

[0228] Step S34: Fine-tune the initial positioning coordinates through ASTCN (Adaptive Space-Time Calibration Network), and the calculation formula is:

[0229]

[0230] where, is the initial position calculated by the four-point spatio-temporal weighted optimization method; is the final event position; is the calibration offset, calculated by the dynamic calibration function , as follows:

[0231]

[0232] wherein, is the feature vector generated by the multi-scale spatio-temporal correction, and is a structured representation compressed, aggregated, and reduced in spatial (sensor distribution) and frequency (scale) dimensions; is the real-time monitored ambient noise level.

[0233] It is achieved through a lightweight neural network and dynamically adjusts the positioning result according to environmental parameters to ensure high robustness under signal attenuation or noise interference.

[0234] After ASTCN calibration, the system outputs the final event position , achieving sub-meter precision positioning.

[0235] Figure 5 is the architecture diagram of the adaptive spatio-temporal calibration network (ASTCN);

[0236] This scheme breaks through the limitations of traditional fixed calibration methods and realizes dynamic adaptive calibration through joint input of multi-scale features and ambient noise levels. It introduces an intelligent optimization mechanism to enable the system to maintain high-precision positioning in scenarios with poor signal quality or strong interference.

[0237] Embodiment 6 is a further improvement based on Embodiment 1, which is as follows:

[0238] The specific steps of preprocessing and standardization in step S03 include:

[0239] Data preprocessing is a basic step for high-precision positioning, aiming to extract effective information from raw vibration data and eliminate interference. The preprocessing process is as follows:

[0240] Wavelet denoising: Wavelet threshold denoising technology is used to process the original vibration signal , effectively separating the target signal from the ambient noise;

[0241] Data standardization: Standardize the vibration intensity to ensure data consistency and facilitate subsequent calculations. The calculation formula is as follows:

[0242] ;

[0243] wherein, is the mean value of the vibration intensity, is the standard deviation, is the normalized vibration intensity data.

[0244] Embodiment 7, this embodiment is a further improvement based on embodiment 1, which is as follows:

[0245] Step S04 specifically includes the following steps:

[0246] To ensure the high consistency of the output results of the sound recognition and vibration positioning branches, the application designs a fusion and verification module to realize the deep cooperation of multi-modal data through probability fusion, conditional probability calculation and weight self-adaptive adjustment. This module takes the improved Bayesian network as the core to realize the cooperative optimization of sound and vibration data through probability fusion. The specific steps are as follows:

[0247] Step S41: Map the event type probability of the sound recognition branch (such as "cable fault: 95%") and the space-time feature of the vibration positioning branch (such as "abnormal vibration intensity at the 500th meter of the optical fiber") to a unified event hypothesis space to form a joint probability distribution, the formula is as follows:

[0248] ;

[0249] The results of the above two branches are input for fusion and verification. In the formula: is the output result of the sound recognition branch (event type and confidence); is the output result of the vibration positioning branch (vibration intensity, position, time, etc.); is the event prior probability; is the normalization factor to ensure the validity of the probability distribution;

[0250] Step S42: Introduce dynamic likelihood ratio (Likelihood Ratio, LR) to quantify the consistency of the output of the sound and vibration branches, identify potential conflicts, and the formula is:

[0251] ;

[0252] In the formula, is the hypothesis that the outputs of the two branches are consistent; is the hypothesis that the outputs of the two branches are inconsistent; is the likelihood ratio, the larger the value, the higher the consistency;

[0253] Step S43: Integrate the probability distributions of the two branches to generate the final fusion result , the formula is:

[0254] ;

[0255] where: is the branch weight, dynamically adjusted according to data quality; is the normalization function, ensuring ;

[0256] The improved Bayesian network fusion mechanism realizes the effective collaboration of multi-source data through event hypothesis space mapping and weighted probability optimization. The dynamic likelihood ratio evaluation method quantifies the consistency between multi-source data by dynamically calculating the likelihood ratio, breaking through the limitations of traditional static threshold.

[0257] Step S44: dynamically adjust the judgment threshold according to the likelihood ratio , the formula is as follows:

[0258] ;

[0259] where, is the basic threshold; is the adjustment coefficient; is the standard deviation of environmental noise;

[0260] Compare the likelihood ratio with the judgment threshold :

[0261] When , it means that the consistency of the fusion result is high enough, and the system accepts the current as the final judgment result without updating;

[0262] When , it means that the consistency is not enough, and the system triggers the secondary analysis mechanism to re-evaluate the data and further update .

[0263] This threshold is used to determine whether to accept the fusion result, directly affecting the decision sensitivity and accuracy of the system. For example, in a noisy environment, the threshold may be appropriately increased to reduce false positives.

[0264] Secondary analysis mechanism principle: when the consistency is lower than the threshold, trigger deep analysis and re-evaluate the data.

[0265] Implementation: combine historical data and contextual features to update an output of a final event probability .

[0266] Effect: this output ensures that the system can still provide reliable judgment results under complex or uncertain conditions, enhancing the robustness of the system.

[0267] Introduce environmental adaptive factor , realize the dynamic optimization of fusion strategy. The system can adjust the parameters according to the real-time environmental changes, greatly improve the robustness and high-precision performance under variable conditions.

[0268] Embodiment 8, this embodiment is a further improvement based on embodiment 7, which is as follows:

[0269] Step S41 also includes an adaptive Bayesian fusion network (ABFN) that dynamically optimizes the fusion process through an environmental adaptive factor, as follows:

[0270] ;

[0271] In the formula, is the environmental adaptive factor, defined as:

[0272] ;

[0273] In the formula, is the environmental feature vector (including noise level, signal attenuation rate, etc.); is a parameter optimized by machine learning.

[0274] Make the system can adjust the fusion strategy according to the real-time environmental changes, ensure high robustness; original quantitative consistency technology, improve the reliability of the system.

[0275] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A dual-branch acoustic vibration fusion event recognition and positioning method based on DAS and AI, characterized by: The steps include: Step S01: Collect sound data from the distributed optical fiber sensing DAS system; capture multi-scale features including time domain features and frequency domain features from the sound data, perform weighted and adaptive fusion on the time domain features and the frequency domain features to obtain a joint feature ; Step S02: From the joint features Extracting high-level spatial features , the Bi-LSTM module is used to perform time series modeling on the feature sequence output by CNN to generate time series features , dynamically adjust the fusion weights of CNN and the Bi-LSTM to obtain the fusion feature , the fusion features Input the fully connected layer, through The function outputs the event type and its confidence, which is recorded as the output result of the sound recognition branch ; Step S03: Use the DAS system to collect vibration intensity data; pre-process and standardize the vibration intensity data to obtain vibration intensity data ; Based on the four-point spatiotemporal weighted optimization positioning method, calculate the initial positioning coordinates of the target point ; Initial positioning coordinates Make fine adjustments and output the final event position , and get the vibration positioning branch output result ; Step S04: Output the result of the voice recognition branch And the vibration positioning branch output results Perform fusion verification to obtain fusion probability distribution, perform consistency assessment through dynamic likelihood ratio, and calculate likelihood ratio , combining the probability distributions of the two branches to generate the final fusion result ; Setting judgment thresholds , and the likelihood ratio With the judgment threshold Compare and determine the final ; The step S04 includes the following steps: Step S41: Map the event type probability of the sound recognition branch and the spatiotemporal features of the vibration localization branch to a unified event hypothesis space to form a joint probability distribution, as shown in the following formula: ； Where: Output the results for the sound recognition branch; Output results for the vibration positioning branch; is the prior probability of the event; is the normalization factor to ensure the validity of the probability distribution; Step S42: Introduce the dynamic likelihood ratio to quantify the consistency of the sound and vibration branch outputs and identify potential conflicts. The formula is: ； Where, Output consistent hypotheses for both branches; The output of the two branches is inconsistent with the hypothesis; is the likelihood ratio, and a larger value indicates a higher consistency; Step S43: Combine the probability distributions of the two branches to generate the final fusion result , the formula is: ； Where: is the branch weight, which is dynamically adjusted according to the data quality; is the normalization function, ensuring .

2. The dual-branch vibroacoustic fusion event recognition and positioning method based on DAS and AI according to claim 1 is characterized in that: The step S01 specifically includes the following steps: Step S11: extracting multi-scale features simultaneously in the time domain and frequency domain; Step S111: Use sliding windows of different lengths or CNN to extract features of different time spans; the output is expressed as ,in is the scale index; Step S112: extracting features of different frequency resolutions through wavelet transform or STFT; The output is represented as ; Step S12: Calculate the attention weights in the time domain and frequency domain respectively through the dual-branch attention module, and weight the features: Step S121: Time domain features of each scale Calculate attention weights: ； in, is a learnable weight matrix; weighted features: ； in, Represents element-wise multiplication; Step S122: Frequency domain features of each scale Calculate attention weights: ； Weighted features: ； in, Represents element-wise multiplication; Step S13: Adaptively adjust the scale weight according to the input signal characteristics, fuse the multi-scale weighted features, and achieve time-frequency joint optimization: Step S131: weighted sum of the weighted features in the time domain and frequency domain respectively: ； ； in, and is a learnable scale weight that represents the contribution of each scale; Step S132: Integrate the time domain and frequency domain features through the joint attention module to obtain the joint feature : 。 3. The method for identifying and locating dual-branch vibroacoustic fusion events based on DAS and AI according to claim 2, characterized in that: The step S02 specifically includes the following steps: Step S21: CNN from joint features Extracting high-level spatial features , achieved through multi-layer convolution operations, the calculation formula is as follows: ； ； The parameters in the formula are described as follows: : No. The input features of the layer (for the first layer, ); : No. The convolution kernel of the layer is usually a four-dimensional tensor ,in are the height and width of the convolution kernel, is the number of input channels, is the number of output channels; : Convolution operation, calculating the sliding window convolution of the input features and the convolution kernel; : No. The bias term of the layer has the shape ; σ: activation function; : No. The output feature map of the layer has the shape of ,in Determined by the input size, convolution kernel size and stride; Step S22: Bi-LSTM performs time series modeling on the feature sequence output by CNN to generate time series features ; The calculation process is as follows: ； ； ； ； ； ； ； in, and are the forward and backward hidden states respectively; Represents element-wise multiplication; Step S23: Dynamically adjust the fusion weights of CNN and Bi-LSTM according to the spatiotemporal characteristics of the input signal to improve the adaptability of feature fusion; the calculation formula is as follows: ； Among them, the adaptive weight Calculated by the following formula: ； in, and is a learnable parameter; Step S24: Extract features from the CNN middle layer and interact with the Bi-LSTM hidden state to obtain , the calculation formula is as follows: ； in, CNN Layer features, For Bi-LSTM in time The hidden state of Step S25: or Input the fully connected layer of CNN, through The function outputs the event type and its confidence, which is recorded as the output result of the sound recognition branch .

4. A dual-branch vibroacoustic fusion event recognition and positioning method based on DAS and AI according to any one of claims 2 or 3, characterized in that: The step S131 further includes: Use a lightweight neural network to predict the best scale combination based on the global characteristics of the input signal: ； in, are the original input features, is a global pooling operation, is the scale selection probability; use Update scale weights and , to achieve dynamic optimization.

5. The method for identifying and locating dual-branch vibroacoustic fusion events based on DAS and AI according to claim 1, characterized in that: The step S03 specifically includes the following steps: Step S31: pre-process and standardize the vibration intensity data to obtain standardized vibration intensity data ; Step S32: Using the four-point spatiotemporal weighted optimization positioning method, the spatiotemporal information of four adjacent measurement points is used to calculate the position coordinates of the target point through weighted optimization. ; Step S321: Obtain data from four adjacent test points, marked as , the position coordinates of each test point are represent; Step S322: define signal parameters, is the initial propagation speed of the signal; The signal reaches The time of each test point; Step S323: Arrival time of signals from four test points Find the earliest arrival time , the formula is: ； Step S324: For each test point Calculate weighting factors , the calculation formula is as follows: ； Where, For the Signal strength or reliability factor at each test point; is the attenuation coefficient; The signal reaches The time of each test point; is the earliest arrival time of the signals at the four test points; Step S325: Using weighting factors and test point locations , calculate the initial positioning coordinates of the target point , the calculation formula is as follows: ； Where, For the The position coordinates of the test points; is the weighting factor; is the sum of the weighted positions of all test points; is the sum of all weights, used for normalization; is the initial positioning coordinate of the target point; Step S33: normalize the signal Perform wavelet transform and decompose it into multi-scale time-frequency features. The calculation formula is as follows: ； in, is the wavelet basis function, and are scale and displacement parameters, respectively; Step S34: Initial positioning coordinates by ASTCN Fine-tune the calculation formula: ； Where, is the initial position calculated by the four-point spatiotemporal weighted optimization method; is the final event position; To calibrate the offset, the dynamic calibration function The calculation formula is as follows: ； Where, Feature vectors generated for multi-scale spatiotemporal correction; For real-time monitoring of ambient noise levels.

6. The method for identifying and locating dual-branch vibroacoustic fusion events based on DAS and AI according to claim 1, characterized in that: The specific steps of preprocessing and standardization in step S03 include: The original vibration signal is de-noised using wavelet threshold denoising technology. to process; The vibration intensity is standardized and the calculation formula is as follows: ； Where, is the mean value of the vibration intensity, is the standard deviation, is the normalized vibration intensity data.

7. The method for identifying and locating dual-branch vibroacoustic fusion events based on DAS and AI according to claim 1, characterized in that: The step S04 further includes the following steps: Step S44: Dynamically adjust the judgment threshold according to the likelihood ratio , the formula is as follows: ； Where, is the basic threshold; is the adjustment coefficient; is the standard deviation of ambient noise; The likelihood ratio With the judgment threshold For comparison: when , the system accepts the current As the judgment result is final, no update is required; when , the system triggers the secondary analysis mechanism, re-evaluates the data, and further updates .

8. The method for identifying and locating dual-branch vibroacoustic fusion events based on DAS and AI according to claim 7, characterized in that: The step S41 further includes: an adaptive Bayesian fusion network, which dynamically optimizes the fusion process through an environmental adaptive factor, and the formula is as follows: ； Where, is the environmental adaptive factor, defined as: ； Where, is the environmental feature vector; are parameters optimized through machine learning.

Images