Long-distance distributed optical fiber acoustic vibration sensing system fused with AI algorithm

Abstract

This invention discloses a long-distance distributed optical fiber acoustic vibration sensing system integrating AI algorithms, comprising an intelligent sensor network constructed from distributed optical fibers. The system divides the optical fiber into multiple monitoring segments, and within each segment, it filters signal segments characterizing environmental noise levels from historical signals to construct an anchor point library for that segment. It extracts the statistical distribution features of the anchor point samples and constructs a style mapping relationship between the segments and a pre-defined common feature space. During real-time monitoring, when an event to be identified is detected, the system extracts event features and maps them to the common feature space through style mapping to obtain normalized event features. The normalized event features are then input into a pre-trained machine learning classification model to output the event type. This invention solves the problem of reduced event recognition accuracy caused by differences in event signal style under different laying environments in long-distance distributed optical fibers.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and smart sensors, and more specifically, to a long-distance distributed fiber optic acoustic vibration sensing system that integrates AI algorithms. Background Technology

[0002] Distributed fiber optic acoustic vibration sensing technology utilizes the backscattering and interference effects of light in optical fibers to convert weak vibrations at different locations along the fiber into time-varying signals, thereby achieving long-distance, continuous, distributed monitoring and location. In scenarios such as perimeter security, pipeline protection, and inspection of transportation and energy infrastructure, the system not only needs to determine whether a disturbance has occurred, but also needs to distinguish the type of disturbance in order to further assess risks and trigger coordinated responses. Because optical fibers often traverse various laying structures and geological environments, the on-site noise is complex and the operating conditions change over a long period of time; event recognition typically relies on the ability to extract signal features and the discriminative power of classification models.

[0003] In existing technologies, the scheme disclosed in CN104568122B, entitled "A Method for Disturbance Event Identification and Localization in a Distributed Fiber Optic Vibration Sensing System," addresses the problem of high false alarms and missed alarms in harsh environments such as strong winds and heavy rains. It proposes a disturbance determination approach based on threshold peak statistics. When there is no disturbance, it uses a limited-amplitude first-order inertial filter to extract and update background information. After detecting a disturbance, it determines the spatiotemporal range and then uses the EM algorithm to perform Gaussian mixture modeling on the event peaks, extracting parameters such as the expectation and covariance of each Gaussian component. These parameters are then fed into a classifier to complete the disturbance type identification and localization. The main focus of this type of scheme is to improve noise resistance through background extraction and peak distribution modeling, and to achieve reliable detection with relatively low complexity.

[0004] Additionally, the scheme disclosed in CN116304813A, entitled "Open Environment Event Recognition Algorithm for Distributed Fiber Optic Vibration Sensing System," uses a convolutional neural network for event classification and introduces an activation feature correction mechanism for open environments. By calculating the difference between the activation features and the average activation features and fitting a correction function, the model corrects the features during the testing phase, thereby reducing false alarms caused by misclassifying unknown events as known events when unknown event types exist. The main focus of this type of scheme is the ability to distinguish between known and unknown categories, and the control of false alarms in open set scenarios.

[0005] However, the above approach still encounters a more specific problem in the actual engineering deployment of long-distance distributed optical fibers. Along the same optical fiber, the laying medium, backfill density, trench structure, coupling efficiency, and local mechanical resonance conditions often differ between different monitoring sections. The signals generated by the same event in different sections are not simply superimposed with background noise of varying intensities, but rather exhibit different frequency band energy distributions, envelope shapes, and phase jitter characteristics, manifesting as a regionally distinctive style difference, similar to different accents for the same sentence in different regions. This style difference is more of a systematic change in the signal propagation link and coupling conditions on the event response, belonging to the shift at the feature distribution level. Simply subtracting the background, updating the background, or statistically modeling the peaks often fails to bring similar events in different sections back to a consistent feature expression. Similarly, merely modifying the activation features in the open environment dimension is insufficient to guarantee that similar events across sections remain comparable in the feature space. The result is that when a model or discrimination rule is trained or calibrated in one segment and then transferred to another segment, similar events may still be split into different categories, and different categories may be confused, which in turn affects the stability of event recognition and alarm in long-distance scenarios. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a long-distance distributed fiber optic acoustic vibration sensing system that integrates AI algorithms, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A long-distance distributed optical fiber acoustic vibration sensing system integrating AI algorithms includes an intelligent sensor network constructed via distributed optical fibers. The distributed optical fibers perform event recognition in the following manner:

[0009] Receive time-series vibration signals propagating along distributed optical fibers and determine the distance location where the signals occur;

[0010] Based on the distance location, the distributed optical fiber is divided into multiple monitoring segments; within each monitoring segment, signal segments that characterize the environmental noise floor of the corresponding monitoring segment are selected from historical signals to construct an anchor point library for the corresponding monitoring segment;

[0011] Extract the statistical distribution characteristics of the anchor point samples in the anchor point library, and calculate the difference between the statistical distribution characteristics and the standard distribution parameters of the preset common feature space. Based on this, construct the style mapping relationship between the monitored road segment and the preset common feature space.

[0012] The monitored road segment is monitored in real time. When an event to be identified is detected, the features of the event to be identified are extracted, and the features of the event to be identified are mapped to the common feature space using the style mapping relationship to obtain normalized event features.

[0013] The normalized event features are input into a pre-trained machine learning classification model to obtain the output event type.

[0014] Preferably, the process of constructing the anchor point library for monitored road sections specifically includes:

[0015] The signals within the monitored road segment are sliced ​​using a preset time window.

[0016] Calculate the short-time energy value and waveform fluctuation variance of the signal within each time window;

[0017] Select time windows with short-term energy values ​​less than a preset silence threshold and waveform fluctuation variance less than a preset stability threshold as candidate segments.

[0018] Calculate the cross-correlation coefficient between adjacent candidate segments. If the cross-correlation coefficient between a preset number of consecutive candidate segments is greater than a preset similarity threshold, then merge the consecutive candidate segments into an anchor sample and store it in the anchor library.

[0019] Preferably, the process of extracting the statistical distribution characteristics of anchor point samples in the anchor point library includes:

[0020] The anchor point samples are subjected to short-time Fourier transform and Mel filtering to generate a Mel frequency cepstral coefficient matrix.

[0021] Calculate the mean vector of the Mel frequency cepstral coefficient matrix of all anchor samples in the anchor library along the feature dimension, and denote it as the source domain mean vector;

[0022] Calculate the covariance matrix of the Mel frequency cepstral coefficient matrix of all anchor samples in the anchor library in the feature dimension, and denote it as the source domain covariance matrix;

[0023] The source domain mean vector and the source domain covariance matrix are combined as the statistical distribution feature.

[0024] Preferably, the steps for obtaining the standard distribution parameters of the preset common feature space are as follows:

[0025] A specific road segment in the distributed optical fiber whose laying environment meets the preset requirements is selected as a reference road segment;

[0026] Collect clean background noise data of the reference road segment under conditions of no external event interference;

[0027] Extract the Mel frequency cepstral coefficient matrix from the clean background noise data;

[0028] Calculate the mean vector and covariance matrix of the Mel frequency cepstral coefficient matrix of the clean background noise data, and use them as the target domain mean vector and target domain covariance matrix, respectively, and combine them as the standard distribution parameters.

[0029] Preferably, the style mapping relationship is achieved by constructing a style mapping matrix, and the formula for calculating the style mapping matrix is:

[0030] ;

[0031] in, The style mapping matrix is... Let be the source domain covariance matrix. Let be the target domain covariance matrix. This represents the operation of the inverse square root of a matrix. This represents the square root operation of a matrix.

[0032] Preferably, the feature vector corresponding to the normalized event feature is obtained in the following way:

[0033] The time-series vibration signal corresponding to the event to be identified is subjected to short-time Fourier transform and Mel filtering to generate the Mel frequency cepstral coefficient matrix of the event to be identified. The original feature vector is obtained by arithmetic averaging of the Mel frequency cepstral coefficient matrix of the event to be identified in the time dimension. ;

[0034] The source domain mean vector is retrieved based on the monitored road segment to which the event to be identified belongs. With the target domain mean vector and call the style mapping matrix. ;

[0035] Calculate the normalized event feature vector using the following formula. :

[0036] ;

[0037] in, The feature vector corresponding to the normalized event features. The original feature vector, Let be the source domain mean vector. The style mapping matrix is... Let be the mean vector of the target domain.

[0038] Preferably, when an event to be identified is detected, a detection window of a first duration and a reference window of a second duration are set, wherein the second duration is longer than the first duration, and the time position of the reference window is located before the detection window; the arithmetic mean of the signal energy within the detection window is calculated and denoted as the short-term energy mean; the arithmetic mean of the signal energy within the reference window is calculated and denoted as the long-term background mean; the ratio of the short-term energy mean to the long-term background mean is calculated; if the ratio exceeds a preset trigger threshold, it is determined that the event to be identified has been detected, and the corresponding signal segment is extracted.

[0039] Preferably, the training method for the pre-trained machine learning classification model includes:

[0040] Different types of event interference are implemented within the reference road section to obtain time-series vibration signal samples corresponding to different event types; the time-series vibration signal samples are input into a convolutional neural network, and supervised learning training is performed with the event type as the label until convergence.

[0041] Preferably, the method further includes a step of dynamically maintaining the anchor point library:

[0042] Assign a lifetime counter to each anchor sample in the anchor library;

[0043] The lifecycle counter is decremented according to a preset time period;

[0044] When a new anchor sample that meets the criteria is detected, it is added to the anchor library and the lifecycle counter is reset to the initial value.

[0045] When the lifecycle counter of an anchor sample reaches zero, the anchor sample is removed from the anchor database.

[0046] After each anchor point library update, the style mapping relationship is recalculated to adapt to environmental changes.

[0047] Preferably, after generating alarm information including distance and location based on the event type, the method further includes:

[0048] Calculate the coverage area of ​​the event to be identified along the length of the optical fiber;

[0049] If the coverage area is less than the preset point event threshold, an alarm is output directly;

[0050] If the coverage area exceeds the point event threshold, then retrieve the normalized event feature vectors of adjacent monitored road segments at the same time.

[0051] The normalized event feature vectors of the current road segment and adjacent road segments are concatenated and input into a preset large-scale event discrimination model for secondary confirmation. The event type is then corrected based on the result of the secondary confirmation.

[0052] The advantage of this invention over the prior art is that it provides an event recognition method for a long-distance distributed fiber optic acoustic vibration sensing system that integrates AI algorithms. Compared with the prior art, its significant benefits are reflected in solving the problem of decreased recognition rate caused by environmental non-uniformity in long-distance sensing and improving the engineering applicability of the system.

[0053] This invention primarily addresses the challenge of "domain differences" in signal characteristics caused by variations in geological environment and fiber coupling conditions in long-distance monitoring scenarios using distributed fiber optic sensing systems. In practical engineering, the same fiber may traverse various environments such as soft soil, hard road surfaces, or suspended structures, resulting in stylistic differences in the signal waveforms generated by the same intrusion behavior at different locations. This makes it difficult for a single model to maintain a high recognition rate across the entire line. This invention constructs an anchor point library based on the background noise characteristics of road segments and establishes a style mapping relationship between each road segment and a preset common feature space using the idea of ​​statistical distribution alignment. This forces heterogeneous signals modulated by the environment to be mapped to a unified standard feature space. This mechanism eliminates the nonlinear interference of environmental factors on signal characteristics, achieves normalization of signal characteristics across the entire line, and allows the backend to train only a general machine learning model to adapt to all monitored road segments. This greatly improves the system's generalization ability and avoids the enormous workload of fine-tuning the model by collecting data separately for each road segment.

[0054] Furthermore, this invention optimizes the system's operational efficiency and deployment cost in terms of specific feature processing and model construction. By extracting the Mel frequency cepstral coefficients and their second-order statistical properties such as mean and covariance matrix of anchor point samples, and combining this with the construction of a clear linear transformation matrix for feature alignment, this method avoids the high computational consumption caused by complex deep neural network domain adaptation algorithms, ensuring the accuracy of feature mapping while meeting the low latency requirements of real-time monitoring. Simultaneously, since all signals are mapped to a common feature space, the system only needs to use the background noise of standard reference road sections superimposed with standard intrusion samples to generate a training set during the training phase, thus achieving full coverage of the complex environment along the entire route, significantly reducing the manpower cost of on-site destructive experimental sample collection and annotation.

[0055] Furthermore, this invention possesses strong environmental adaptability and false alarm resistance. By introducing a dynamic maintenance mechanism for the anchor point library, the system can automatically update background features based on the lifecycle counter, thereby adapting to changes in environmental noise caused by weather changes such as rainfall and icing, ensuring continuous identification stability. Combined with a secondary confirmation mechanism based on multi-segment collaboration, the system can effectively distinguish between point-like intrusion events such as excavation and large-scale area interference such as vehicle movement by analyzing the distribution width of events and fusing features of adjacent road segments. This significantly reduces the false alarm rate under complex operating conditions and substantially improves the actual protective effectiveness of the system. Attached Figure Description

[0056] Figure 1 This is the overall logic flowchart of the event recognition process of this invention;

[0057] Figure 2 A schematic diagram showing optical fibers crossing different geological environments is presented.

[0058] Figure 3 This is a schematic diagram illustrating the specific process of constructing the monitoring section anchor point database according to the present invention;

[0059] Figure 4 This is a schematic diagram of the style mapping of this invention. Detailed Implementation

[0060] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0061] like Figure 1 As shown, this invention provides a long-distance distributed optical fiber acoustic vibration sensing system integrating AI algorithms, used to identify events and output event type and distance location of vibration signals propagating along distributed optical fibers. In some specific embodiments, the system as a whole can be composed of a distributed optical fiber body, an optical fiber demodulation host, an edge computing unit, a central analysis unit, and an alarm publishing unit. The distributed optical fiber is used to lay in the area to be monitored to form an intelligent sensor network; the optical fiber demodulation host is used to inject probe light into the optical fiber and receive the echo to obtain a vibration time series that changes over time; the edge computing unit is used to complete event triggering, signal preprocessing, and feature extraction; the central analysis unit is used for anchor point library maintenance, style mapping calculation, and machine learning classification; the alarm publishing unit is used to combine event type and distance location to generate alarm information, and can further perform secondary confirmation of events over a large area.

[0062] like Figure 2As shown, distributed optical fibers may traverse different geological environments or laying structures, such as soil layers of varying hardness, clay, rock, and potentially gravel layers, rock strata, pipe galleries, cable trays, or buried conduits. Different environments lead to significant differences in coupling efficiency, noise floor spectral morphology, and attenuation characteristics, resulting in different characteristic distributions of similar events across different road segments. To ensure the classification model remains stable over long distances and in multiple environments, this invention introduces an anchor point library and style mapping at the monitored road segment level, unifying the characteristics of each road segment into a common feature space before making classification decisions.

[0063] In some embodiments, the process of receiving time-series vibration signals propagating along a distributed optical fiber and determining the distance position can be achieved using the conventional positioning mechanism of a distributed acoustic sensing (DAS) system. The fiber optic demodulation host emits pulsed light or coherent probe light at a preset repetition frequency, and obtains the distance coordinates along the fiber length direction by using the arrival time of the echo signal and system calibration. The distance resolution can be determined by the sampling rate, pulse width, and signal processing method, typically ranging from 0.5m to 10m. To ensure stable distance position, calibration can be performed once or multiple times during system deployment, writing the actual fiber laying length, light speed conversion factor, and host internal delay compensation into a configuration table; when temperature changes significantly, recalibration can be performed daily or weekly to reduce drift.

[0064] In another embodiment, the distance position can also be established by tapping a known landmark or by artificially triggering an event to establish a distance anchor point, and then perform linear or piecewise linear correction on the entire distance axis to adapt to the non-ideal length mapping caused by local slack or coiling of the optical fiber.

[0065] The process of dividing a distributed optical fiber into multiple monitoring segments based on distance and location allows for segment length settings to be configured according to engineering requirements. A common approach is to divide the entire length into segments at fixed distances, such as 50m, 100m, or 200m per segment. When environmental changes are significant, non-uniform segmentation can be used to ensure each segment is under relatively consistent geological or structural conditions. The purpose of segmentation is to ensure each segment has a relatively stable background noise pattern, facilitating the construction of a dedicated anchor point library for that segment and the generation of reliable style mapping relationships.

[0066] like Figure 3 As shown, within each monitored road segment, signal segments characterizing environmental noise background are selected from historical signals to construct an anchor point library. The anchor point library can store multiple anchor point samples, each corresponding to a low-energy, stable, and similar background noise segment over a given period. The anchor point library can be established during the learning phase after initial system deployment or continuously supplemented during operation.

[0067] In one embodiment, signal slicing uses a preset time window to slide-segment the signal within a road segment. The time window length can be from 0.2s to 5s, and the sliding step size can be from 10% to 50% of the time window length to achieve a balance between computational complexity and stability. Subsequently, the short-time energy value and waveform fluctuation variance within each time window are calculated. The short-time energy value can be obtained from the sum of squares or root mean square of the sampling points within the window; the waveform fluctuation variance can be obtained from the variance of the signal amplitude within the window or the variance after first-order differencing, used to reflect whether the window is stable. The silence threshold and stability threshold can be adaptively selected based on the historical statistics of the road segment. For example, the 10th percentile of the short-time energy distribution over the past 24 hours can be used as the silence threshold, and the 20th percentile of the fluctuation variance distribution can be used as the stability threshold. The reason for using quantiles is that the absolute noise intensity of different road segments may vary significantly; using quantiles allows the screening rules to automatically scale with the road segment's noise floor level, reducing manual parameter tuning.

[0068] After selecting time windows with short-term energy values ​​below the silence threshold and waveform fluctuation variance below the stationarity threshold as candidate segments, the cross-correlation coefficient between adjacent candidate segments is further calculated. The cross-correlation coefficient can be calculated using normalized cross-correlation, with a value range of -1 to 1. The similarity threshold can be set between 0.7 and 0.98, and the consecutive number can be set between 3 and 20. The consecutive number constraint is introduced to avoid occasional short-term silence segments being mistakenly included as anchor points; only stable background noise that consistently exhibits similarity over a period of time is considered representative of the background noise style of that road segment. When the cross-correlation coefficients between a predetermined number of consecutive candidate segments are all greater than the similarity threshold, these consecutive candidate segments are merged into one anchor point sample and stored in the anchor point library. After merging, information such as the start and end times, road segment number, average energy, and spectral summary of the anchor point sample can be recorded for subsequent management.

[0069] In other embodiments, candidate fragment selection can be replaced by stability determination based on spectral entropy, zero crossover rate, or bandpass energy ratio; cross-correlation similarity can be replaced by cosine similarity, dynamic time-warped distance, or distance metric based on MFCC features. The aim is to improve compatibility with different devices and sampling rate configurations while ensuring that the anchor point samples are indeed background noise.

[0070] After constructing the anchor point library, it is necessary to extract the statistical distribution characteristics of the anchor point samples. In one embodiment, the anchor point samples are subjected to short-time Fourier transform and Mel filtering to generate a Mel frequency cepstral coefficient matrix. The frame length of the short-time Fourier transform can be 20ms to 40ms, and the frame shift can be 10ms to 20ms; the number of Mel filter banks can be 40 to 128; and the dimension of the Mel frequency cepstral coefficients can be 13 to 40. The frame length and frame shift determine the trade-off between time resolution and frequency resolution, while the number of Mel filter banks and the cepstral dimension determine the feature representation capability and computational cost. To adapt to long-distance real-time systems, a common configuration is a frame length of 25ms, a frame shift of 10ms, 64 Mel filter banks, and a cepstral dimension of 20.

[0071] After obtaining the Mel frequency cepstral coefficient matrix of the anchor point samples, the mean vector of all anchor point samples in the anchor point library is calculated along the feature dimension as the source domain mean vector, and the covariance matrix is ​​calculated as the source domain covariance matrix. The mean vector characterizes the central position of the road segment's noise floor along the common feature dimension, while the covariance matrix characterizes the fluctuation scale of the road segment's noise floor along each feature dimension and the correlation between different dimensions. Combining these two characteristics serves as the statistical distribution feature of the road segment's anchor point library, allowing for the subsequent construction of a style mapping relationship from the road segment to the common feature space based on the distribution differences.

[0072] like Figure 4 As shown, the standard distribution parameters of the common feature space can be obtained by selecting a reference road segment. The reference road segment can be a specific road segment whose paving environment meets the preset standard soil requirements and whose coupling efficiency is satisfactory. The standard soil requirements can be provided by the engineering party, such as meeting backfill compaction standards, meeting design depth requirements, and having no continuous mechanical vibration sources nearby. The coupling efficiency can be evaluated by the signal-to-noise ratio of background noise or the amplitude of the test event response. Of course, the reference road segment can also be selected manually, or even one segment can be manually designated as the reference road segment from all road segments.

[0073] After collecting clean background noise data of the reference road segment under conditions free from external interference, a Mel-frequency cepstral coefficient matrix is ​​generated using the same feature extraction process as the anchor points. Its mean vector and covariance matrix are then calculated and used as the mean vector and covariance matrix of the target domain, respectively. The combination of these two matrices constitutes the standard distribution parameters of the common feature space. The purpose of using the reference road segment is to establish a uniform scale that is repeatable and interpretable in engineering, aligning different road segments to the same standard. This allows subsequent classification models to learn only the event discrimination rules within this standard space.

[0074] After obtaining the source domain covariance matrix and the target domain covariance matrix, the style mapping relationship can be realized through the style mapping matrix. The style mapping matrix is ​​calculated using the formula... ,in, For style mapping matrix, Let be the source domain covariance matrix. Let the target domain covariance matrix be... This represents the operation of the inverse square root of a matrix. This represents the square root operation of a matrix. Specifically, the inverse square root operation is used to decorrelate and scale-normalize the features of the source domain, essentially whitening the source domain distribution. The square root operation is used to recolor the whitened features onto the scale and correlation structure of the target domain. After this processing, the features before and after mapping are statistically closer to the target domain distribution. The square root and inverse square root of a matrix can be implemented using eigenvalue decomposition or singular value decomposition. The covariance matrix is ​​decomposed into an eigenvector matrix and an eigenvalue diagonal matrix, and then the square root or the reciprocal of the square root is taken. For numerical stability, a very small regularization term, on the order of 1e-6 to 1e-3, can be added to the diagonal of the covariance matrix to prevent the operation from diverging due to excessively small eigenvalues.

[0075] The process of real-time monitoring and detection of events to be identified can be triggered based on the ratio of short-term energy to long-term background. In one embodiment, a detection window of a first duration and a reference window of a second duration are set, with the second duration being longer than the first duration and the reference window preceding the detection window. The first duration can be 0.1s to 2s for sensitively capturing sudden events; the second duration can be 2s to 60s for stable background estimation. The arithmetic mean of the signal energy within the detection window is calculated as the short-term energy mean, and the arithmetic mean of the signal energy within the reference window is calculated as the long-term background mean. Then, the ratio between the two is calculated. If the ratio exceeds the trigger threshold, it is determined that an event to be identified has been detected, and the corresponding signal segment is extracted. The trigger threshold can be 1.5 to 10, and the specific value can be determined according to the target false alarm rate. If the focus in the engineering field is on preventing missed alarms, the trigger threshold can be lowered and the subsequent classification model can be used to suppress false alarms; if the focus is on reducing alarm interference, the trigger threshold can be increased and the first duration extended to enhance trigger robustness. In another embodiment, the trigger determination may also incorporate a duration constraint, such as triggering only if the ratio continuously exceeds a threshold for more than 0.2 seconds, in order to suppress instantaneous spike noise.

[0076] Once an event to be identified is detected and a signal segment is captured, event features need to be extracted and mapped to a common feature space. In one embodiment, a short-time Fourier transform and Mel filtering are performed on the event signal segment to generate the Mel frequency cepstral coefficient matrix of the event, and the original feature vector is obtained by arithmetic averaging over the time dimension. The purpose of time-dimensional averaging is to compress variable-length event segments into fixed-length vectors, which facilitates subsequent mapping and classification. If it is necessary to retain the time structure, another embodiment can also use time-dimensional segmented averaging or fixed-frame-count truncation to obtain fixed-length sequence features, which are then passed to the sequence model for classification.

[0077] Based on the monitored road segment to which the event belongs, the source domain mean vector and the common space target domain mean vector of that road segment are retrieved, and the style mapping matrix is ​​also invoked. According to the formula Calculate the normalized event feature vector. The formula means: first use... The event characteristics are shifted from the source domain center to the zero center, and then... Complete the transformation from the source domain scale and correlation structure to the target domain scale and correlation structure, and finally add... The features are shifted to the center of the target domain. This normalization makes similar events from different road segments more likely to fall into similar regions, significantly reducing the difficulty for classification models to simultaneously fit multiple noise levels. In the above formula, the subscripts... S Indicates the source domain; subscript T Indicates the target domain, subscript raw Indicates the initial state, subscript norm This represents the normalized state after alignment.

[0078] In another embodiment, style transfer can be achieved without matrix formula calculation. Instead, feature alignment can be accomplished by training an additional learnable style transfer model. The role of this style transfer model is to receive the original event features or time-frequency features of a certain monitored road segment and output normalized features that fall within the common feature space distribution, so that the subsequent classification model still only needs to deal with data under a uniform distribution.

[0079] One implementation involves configuring a lightweight feature transformation network as a road segment adapter for each monitored road segment. The network structure can be a multilayer perceptron or a 1D convolutional network. The input to the multilayer perceptron is the original feature vector, which undergoes 2 to 4 layers of linear transformation and nonlinear activation to output a normalized feature vector of the same dimension. If the input is a Mel-frequency cepstral coefficient matrix or a Mel-frequency spectrum, a shallow convolutional network can be used to convolutionally encode the time-frequency map before mapping it to a common feature space. The parameters of the road segment adapter are obtained through optimization algorithms, without needing to explicitly calculate the source domain covariance matrix and the target domain covariance matrix, nor perform inverse square root and square root operations.

[0080] The goal of training this style transfer model is to transform the background anchor point features of a monitored road segment so that they statistically approximate the background distribution of the reference road segment. Specifically, anchor point samples are extracted from the anchor point library of the monitored road segment, and their feature representations are extracted. Simultaneously, background samples are extracted from the clean background noise data of the reference road segment, and their corresponding feature representations are extracted. The anchor point features of the monitored road segment are input into the road segment adapter to obtain the transformed features, and a distribution alignment loss is constructed as the training objective. The distribution alignment loss can be a combination of mean difference and covariance difference, making the mean of the transformed features converge towards the mean of the reference road segment, and the correlation structure between feature dimensions converge towards the reference road segment; alternatively, kernel mean matching loss or adversarial discrimination loss can be used to make it difficult for the discriminator to distinguish the transformed features from the features of the reference road segment, thus approximating the distribution of the reference road segment. Parameter optimization can use Adam or SGD, with a learning rate of 1e-4 to 1e-2. Training stops when the distribution difference of the validation set tends to stabilize, and the convergence criterion is that the distribution alignment loss decreases less than a preset threshold within 10 consecutive rounds.

[0081] When the system is online and running, after detecting an event to be identified and extracting its original features, style mapping can be performed without calling formulas. Instead, the corresponding road segment adapter is invoked based on the monitored road segment to which the event belongs, directly converting the event features into normalized event features. These normalized features are then input into a pre-trained classification model to output the event type. To adapt to environmental changes, the road segment adapter can also be retrained or incrementally updated using a dynamic maintenance mechanism of the anchor point library. For example, when the anchor point library update ratio exceeds a preset threshold, the road segment adapter is fine-tuned several times using the latest anchor point samples and reference road segment background samples, ensuring that the conversion relationship maintains alignment as the background noise changes. In this way, style conversion from monitored road segments to a common feature space can be achieved through model parameter optimization without relying on explicit formula calculations.

[0082] Furthermore, normalized event features are obtained, and then these features are input into a pre-trained machine learning classification model to output the event type. The classification model can be implemented using a convolutional neural network, with the normalized feature vector as input. It can employ a multi-layer fully connected network or a 1D convolutional network. To adapt to edge device inference, a lightweight structure can be chosen, such as 3 to 6 convolutional layers, with the number of channels increasing layer by layer from 16 to 128, combined with batch normalization and ReLU activation. Finally, global average pooling and fully connected layers are used to output the event category probability. Event categories can be defined according to the application as multiple intrusion events, such as digging, knocking, vehicle running over, footsteps, leak impact, or other categories required by engineering.

[0083] In one embodiment, the machine learning classification model completes data collection and training on a reference road segment. Different types of event disturbances are implemented within the reference road segment to form labeled data samples. These disturbances can include forms of disturbance related to the engineering scenario, such as knocking, digging, trampling, vehicle running over, and dragging friction. Each time an event disturbance is implemented, a distributed fiber optic acoustic vibration sensing system continuously collects the corresponding time-series vibration signal and records the start and end times of the event and the event type label, obtaining time-series vibration signal samples corresponding to different event types. To reduce training bias caused by amplitude differences between samples, amplitude or energy normalization can be performed on each segment of the time-series vibration signal sample. Samples can also be truncated or padded to a preset length to ensure the input length meets the network training requirements. The preset length can be between 0.5s and 10s; shorter lengths are used for transient knocking events, and longer lengths are used for continuous digging or running over events, depending on the event's duration characteristics. During the training phase, the time-series vibration signal samples are input into a convolutional neural network, and supervised learning training is performed using the event type as a label until convergence. During supervised learning training, the loss function is cross-entropy loss, and the optimizer is Adam or SGD. The initial learning rate can be between 1e-4 and 1e-2, and it adaptively decays according to the performance on the validation set. During training, the training set and validation set are divided according to a preset ratio, such as 8:2. The validation set accuracy or validation set loss is used as the convergence criterion. When the validation set metric improves by less than 0.1% within 10 consecutive rounds or the validation set loss decreases by less than a preset threshold, convergence is determined and training stops, resulting in a machine learning classification model. To enhance the model's adaptability to background fluctuations at different collection times for reference road sections, background segments collected at different times can be mixed into the training samples, or small-amplitude random noise and random amplitude scaling can be superimposed on the time series at the input, allowing the model to learn more stable discriminative features without changing the event type labels.

[0084] The anchor point library needs to be dynamically maintained as the environment changes. In one embodiment, a lifespan counter is assigned to each anchor point sample in the library, and the counter is decremented at a preset time period, for example, once every 1 hour. The initial value can be 24 to 720, indicating that the anchor point can be retained for 1 to 30 days. When a new anchor point sample that meets the conditions is detected, it is added to the anchor point library and its counter is reset to the initial value. At the same time, similar anchor points can be merged to remove redundancy according to a strategy to control the size of the anchor point library, for example, retaining 50 to 500 anchor points per road segment. When the counter of an anchor point sample reaches zero, it is removed from the anchor point library to avoid the influence of outdated background noise features on the current mapping. After each update of the anchor point library, the source domain mean vector and source domain covariance matrix are recalculated, and the style mapping matrix is ​​updated accordingly, so that the mapping relationship automatically tracks the slow change of background noise. In order to reduce the computational overhead caused by frequent updates, update trigger conditions can be set, such as recalculating only when the cumulative number of new or deleted anchor points exceeds 5% of the total; or an incremental update method with sliding statistics can be used to avoid full traversal.

[0085] When generating alarm information, the system can output a combination of event type and distance location, and further calculate the coverage area of ​​the event along the fiber optic length. The coverage area can be obtained by statistically analyzing the lengths of consecutive distance cells exceeding a trigger threshold or a certain energy threshold on the distance axis; the threshold for point events can range from 2m to 50m. If the coverage area is smaller than the point event threshold, it typically corresponds to single-point impact, localized excavation, etc., and the system can directly output an alarm. If the coverage area exceeds the point event threshold, it may correspond to long-distance vehicle crushing, pipeline resonance propagation, or large-scale construction; in this case, classification using only a single road segment may lead to misjudgment. Therefore, the system can retrieve the normalized event feature vectors of adjacent monitored road segments at the same time, concatenate the feature vectors of the current road segment and adjacent road segments, input them into a preset large-scale event discrimination model for secondary confirmation, and correct the event type based on the secondary confirmation result. The large-scale event discrimination model can use a lightweight multilayer perceptron or an attention-based sequence model, using features from multiple adjacent road segments as sequence input, thereby reducing misjudgment by utilizing spatial consistency constraints. The training data for secondary confirmation can be formed by aligning samples of the same type of large-scale events across multiple road segments and introducing samples with different lengths of coverage to enhance the model, enabling it to learn to distinguish between real large-scale events and pseudo-large-scale phenomena caused by propagation or accidental superposition.

[0086] To ensure the system is easy to deploy and maintain on-site, an automated strategy can be provided for the selection of thresholds and parameters. Default ranges can be initially given for the silent threshold, stability threshold, similarity threshold, and trigger threshold, allowing the system to adaptively converge based on historical statistics for each road segment during the initial learning phase. The default ranges can be set as follows for rapid startup: silent threshold from 5% to 20% of the energy quantile, stability threshold from 10% to 30% of the variance quantile, similarity threshold from 0.8 to 0.95, trigger threshold from 2 to 6, and point event threshold from 5m to 20m. If there is strong noise and many false alarms on-site, the trigger threshold and similarity threshold can be increased; if there are weak events and many missed alarms, the trigger threshold can be decreased and the first duration shortened, while simultaneously increasing the anchor point library size to improve style mapping accuracy.

[0087] Through the above implementation methods, this invention characterizes the background noise style of each monitored road segment using an anchor point library under long-distance, multi-geological, and multi-coupling conditions, and unifies event features into a common feature space through style mapping. Then, a pre-trained classification model completes the event type identification. At the same time, with the dynamic maintenance of the anchor point library and a large-scale event secondary confirmation mechanism, the system maintains stable identification performance and an interpretable processing link under changes in the actual engineering environment.

[0088] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A long-distance distributed fiber optic acoustic vibration sensing system integrating AI algorithms, characterized in that, This includes an intelligent sensor network, which is constructed using distributed optical fibers, and the distributed optical fibers perform event recognition in the following manner: Receive time-series vibration signals propagating along distributed optical fibers and determine the distance location where the signals occur; Based on the distance location, the distributed optical fiber is divided into multiple monitoring segments; within each monitoring segment, signal segments that characterize the environmental noise floor of the corresponding monitoring segment are selected from historical signals to construct an anchor point library for the corresponding monitoring segment; Extract the statistical distribution characteristics of the anchor point samples in the anchor point library, and calculate the difference between the statistical distribution characteristics and the standard distribution parameters of the preset common feature space. Based on this, construct the style mapping relationship between the monitored road segment and the preset common feature space. The monitored road segment is monitored in real time. When an event to be identified is detected, the features of the event to be identified are extracted, and the features of the event to be identified are mapped to the common feature space using the style mapping relationship to obtain normalized event features. The normalized event features are input into a pre-trained machine learning classification model to obtain the output event type; The process of constructing an anchor point database for monitored road sections specifically includes: The signals within the monitored road segment are sliced ​​using a preset time window. Calculate the short-time energy value and waveform fluctuation variance of the signal within each time window; Select time windows with short-term energy values ​​less than a preset silence threshold and waveform fluctuation variance less than a preset stability threshold as candidate segments. Calculate the cross-correlation number between adjacent candidate segments. If the cross-correlation number between a preset number of consecutive candidate segments is greater than a preset similarity threshold, then merge the consecutive candidate segments into an anchor sample and store it in the anchor library. The process of extracting the statistical distribution characteristics of anchor point samples in the anchor point library includes: The anchor point samples are subjected to short-time Fourier transform and Mel filtering to generate a Mel frequency cepstral coefficient matrix. Calculate the mean vector of the Mel frequency cepstral coefficient matrix of all anchor samples in the anchor library along the feature dimension, and denote it as the source domain mean vector; Calculate the covariance matrix of the Mel frequency cepstral coefficient matrix of all anchor samples in the anchor library in the feature dimension, and denote it as the source domain covariance matrix; The source domain mean vector and the source domain covariance matrix are combined as the statistical distribution feature; The steps for obtaining the standard distribution parameters of the preset common feature space are as follows: A specific road segment in the distributed optical fiber whose laying environment meets the preset requirements is selected as a reference road segment; Collect clean background noise data of the reference road segment under conditions of no external event interference; Extract the Mel frequency cepstral coefficient matrix from the clean background noise data; Calculate the mean vector and covariance matrix of the Mel frequency cepstral coefficient matrix of the clean background noise data, and use them as the target domain mean vector and target domain covariance matrix, respectively, and combine them as the standard distribution parameters; The style mapping relationship is achieved by constructing a style mapping matrix, and the formula for calculating the style mapping matrix is ​​as follows: ； in, The style mapping matrix is... Let be the source domain covariance matrix. Let be the target domain covariance matrix. This represents the operation of the inverse square root of a matrix. Represents the square root operation of a matrix; The feature vector corresponding to the normalized event feature is obtained in the following way: The time-series vibration signal corresponding to the event to be identified is subjected to short-time Fourier transform and Mel filtering to generate the Mel frequency cepstral coefficient matrix of the event to be identified. The original feature vector is obtained by arithmetically averaging the Mel frequency cepstral coefficient matrix of the event to be identified in the time dimension. ; The source domain mean vector is retrieved based on the monitored road segment to which the event to be identified belongs. With the target domain mean vector and call the style mapping matrix. ; Calculate the normalized event feature vector using the following formula. : ； in, The feature vector corresponding to the normalized event features. The original feature vector, Let be the source domain mean vector. The style mapping matrix is... The mean vector of the target domain; When an event to be identified is detected, a detection window of a first duration and a reference window of a second duration are set, wherein the second duration is longer than the first duration and the time position of the reference window is before the detection window; the arithmetic mean of the signal energy within the detection window is calculated and denoted as the short-term energy mean; the arithmetic mean of the signal energy within the reference window is calculated and denoted as the long-term background mean; the ratio of the short-term energy mean to the long-term background mean is calculated; if the ratio exceeds a preset trigger threshold, it is determined that the event to be identified has been detected, and the corresponding signal segment is extracted; It also includes the step of dynamically maintaining the anchor point library: Assign a lifetime counter to each anchor sample in the anchor library; The lifecycle counter is decremented according to a preset time period; When a new anchor sample that meets the criteria is detected, it is added to the anchor library and the lifecycle counter is reset to the initial value. When the lifecycle counter of an anchor sample reaches zero, the anchor sample is removed from the anchor database. After each update of the anchor point library, the style mapping relationship is recalculated to adapt to environmental changes; After generating alarm information including distance and location based on the event type, the process further includes: Calculate the coverage area of ​​the event to be identified along the length of the optical fiber; If the coverage area is less than the preset point event threshold, an alarm is output directly; If the coverage area exceeds the point event threshold, then retrieve the normalized event feature vectors of adjacent monitored road segments at the same time. The normalized event feature vectors of the current road segment and adjacent road segments are concatenated and input into a preset large-scale event discrimination model for secondary confirmation. The event type is then corrected based on the result of the secondary confirmation.

2. The system according to claim 1, characterized in that, The training method for the pre-trained machine learning classification model includes: Different types of event interference are implemented within the reference road section to obtain time-series vibration signal samples corresponding to different event types; the time-series vibration signal samples are input into a convolutional neural network, and supervised learning training is performed with the event type as the label until convergence.

Images