A distributed optical fiber vibration event identification and early warning method, system and electronic equipment

Abstract

The application provides a distributed optical fiber vibration event identification and early warning method and system and electronic equipment, and relates to the technical field of distributed optical fiber acoustic sensing. The method comprises the following steps: arranging a matrix according to vibration data to obtain a two-dimensional waterfall diagram; performing double-criterion denoising and feature extraction processing on the two-dimensional waterfall diagram to obtain physical statistical features; constructing multiple space-time blocks according to the physical statistical features to obtain a space-time block feature vector corresponding to each space-time block; predicting the space-time block feature vector by using an incremental discriminant learning classification model to obtain a block-level identification result; performing decision correction on each block-level identification result to obtain a corrected block-level identification result, determining an effective event segment based on all the corrected block-level identification results, and obtaining a warning result. The application improves the identification and early warning accuracy of distributed optical fiber vibration events.

Description

Technical Field

[0001] This invention relates to the field of distributed fiber optic acoustic sensing technology, and more specifically, to a distributed fiber optic vibration event identification and early warning method, system, and electronic device. Background Technology

[0002] Distributed optical fiber, with its advantages of long-distance and continuous distributed monitoring, is widely used in safety monitoring of oil and gas pipelines, pipe corridors, and railway perimeters. Since abnormal disturbances such as human construction, mechanical excavation, and impacts from foreign objects can easily cause safety accidents such as facility damage, leaks, and line interruptions, it is necessary to identify various vibration events along the distributed optical fiber route and issue timely warnings.

[0003] In related technologies, deep learning or traditional machine learning are mainly used to identify and warn of vibration events in distributed optical fibers. Deep learning methods convert waterfall plots into images or time-series signals for processing; however, these models have a large number of parameters, long training cycles, and high deployment barriers. Running on edge devices with limited computing power results in significant latency, failing to guarantee real-time warnings. Meanwhile, traditional machine learning is relatively simple and crude in feature extraction, model structure design, and training strategies, leading to poor sample recognition capabilities. Therefore, both methods are prone to false or missed warnings during vibration event identification and early warning, thus failing to meet the requirements of on-site security monitoring. Summary of the Invention

[0004] The problem addressed by this invention is how to improve the accuracy of identification and early warning of distributed optical fiber vibration events.

[0005] To address the above problems, this invention provides a distributed optical fiber vibration event identification and early warning method, system, and electronic device.

[0006] In a first aspect, the present invention provides a distributed optical fiber vibration event identification and early warning method, comprising: Vibration data of distributed optical fibers are acquired in real time, and the vibration data is arranged in a matrix to obtain a two-dimensional waterfall plot of the vibration data. The two-dimensional waterfall plot is subjected to dual-criteria denoising and feature extraction to obtain the physical statistical features of the two-dimensional waterfall plot; Multiple spatiotemporal blocks are constructed based on the physical statistical characteristics, and a spatiotemporal block feature vector corresponding to each spatiotemporal block is obtained; By using a preset incremental discriminative learning classification model, prediction is made based on the spatiotemporal block feature vectors to obtain the block-level recognition result of each spatiotemporal block feature vector; Each block-level recognition result is then corrected to obtain the corrected block-level recognition result. Based on all the corrected block-level identification results, the valid event segments of the distributed optical fiber are determined; Based on the effective event segments, vehicle diagonal texture filtering is performed to obtain the warning results for the effective event segments.

[0007] Optionally, the real-time acquisition of vibration data from the distributed optical fiber, and the matrix arrangement of the vibration data to obtain a two-dimensional waterfall plot of the vibration data, includes: Vibration data of the distributed optical fiber are acquired using a distributed optical fiber acoustic sensing device at a preset acquisition frequency. The distributed optical fiber is divided into multiple spatial channels according to a preset spatial resolution; Based on the vibration data, determine the vibration response amplitude of each of the spatial channels at each acquisition time; Based on the vibration response amplitude of each spatial channel at each acquisition time, a two-dimensional vibration matrix is ​​obtained by arranging the data matrix according to the time dimension and the spatial channel dimension. The two-dimensional vibration matrix is ​​used as the two-dimensional waterfall diagram.

[0008] Optionally, the step of performing dual-criterion denoising and feature extraction on the two-dimensional waterfall plot to obtain the physical statistical features of the two-dimensional waterfall plot includes: Based on the vibration response amplitude of each spatial channel in the two-dimensional waterfall diagram, determine the low amplitude ratio and high amplitude point count for each spatial channel; Based on the comparison results of the low amplitude ratio of the spatial channel with the preset low amplitude ratio threshold, and the comparison results of the high amplitude point count of the spatial channel with the preset high amplitude point quantity threshold, denoising processing is performed to obtain a denoised two-dimensional waterfall image. The denoised two-dimensional waterfall plot is extracted by sliding according to a time window of a preset size to obtain the short-time energy, high amplitude ratio, number of high amplitude continuous segments, root mean square of short-time energy, peak value of vibration response, and excess value of vibration response kurtosis for each spatial channel within each time window. The short-time energy, high amplitude ratio, number of continuous high amplitude segments, root mean square of short-time energy, peak value of vibration response, and excess value of vibration response kurtosis of all spatial channels within all time windows are used as the physical statistical characteristics of the two-dimensional waterfall plot.

[0009] Optionally, the step of constructing multiple spatiotemporal blocks based on the physical statistical characteristics to obtain a spatiotemporal block feature vector corresponding to each spatiotemporal block includes: The clipping window of the spatiotemporal block is determined based on the preset feature frame domain in the time dimension and the preset neighborhood range in the spatial channel dimension. Based on the clipping window, multiple spatiotemporal blocks are extracted from the physical statistical features; The spatial channels in the neighborhood of the spatiotemporal block are spliced ​​together in a preset order according to the short-time energy, high amplitude ratio, number of high amplitude continuous segments, root mean square of short-time energy, peak value of vibration response, and excess value of vibration response kurtosis in all time windows of the feature frame domain corresponding to the spatiotemporal block, to obtain the spatiotemporal block feature vector corresponding to the spatiotemporal block.

[0010] Optionally, the step of predicting based on the spatiotemporal block feature vectors using a preset incremental discriminative learning classification model to obtain the block-level recognition result for each spatiotemporal block feature vector includes: The spatiotemporal block feature vector is input into the incremental discriminant learning classification model; The probability value of the spatiotemporal block feature vector under a preset category is obtained by calculating the probability value of the spatiotemporal block feature vector based on the hidden mapping layer and output layer of the incremental discriminative learning classification model, according to the spatiotemporal block feature vector. The preset category includes background class, artificial disturbance class and mechanical disturbance class. By comparing the probability values ​​corresponding to the background class, the artificial disturbance class, and the mechanical disturbance class, the disturbance category corresponding to the spatiotemporal block feature vector is determined, and the disturbance category is used as the block-level recognition result of the spatiotemporal block feature vector.

[0011] Optionally, the step of making decision corrections for each block-level recognition result to obtain corrected block-level recognition results includes: When the perturbation category of the spatiotemporal block feature vector is the artificial perturbation class, if the probability value of the spatiotemporal block feature vector in the artificial perturbation class is less than the preset minimum artificial confidence level, then the perturbation category of the spatiotemporal block feature vector is corrected to the background class, and the corrected block-level recognition result is obtained. When the perturbation category of the spatiotemporal block feature vector is the mechanical perturbation category, if the probability value of the spatiotemporal block feature vector in the mechanical perturbation category is less than the preset minimum confidence threshold for machinery, then the perturbation category of the spatiotemporal block feature vector is corrected to the background category, and the corrected block-level recognition result is obtained. When the probability value of the spatiotemporal block feature vector in the mechanical disturbance class is greater than or equal to a preset mechanical forced reclassification threshold, and the difference between the probability value of the spatiotemporal block feature vector in the mechanical disturbance class and the probability value of the spatiotemporal block feature vector in the background class is greater than or equal to a preset mechanical relative advantage margin, and the probability value of the spatiotemporal block feature vector in the mechanical disturbance class is greater than or equal to the probability value of the spatiotemporal block feature vector in the artificial disturbance class, then the disturbance category of the spatiotemporal block feature vector is corrected to the mechanical disturbance class, and the corrected block-level recognition result is obtained.

[0012] Optionally, determining the valid event segments of the distributed optical fiber based on all the corrected block-level identification results includes: Based on the corrected block-level identification results of all the spatiotemporal block feature vectors, the disturbance category of the spatial channel contained in the spatiotemporal block corresponding to each spatiotemporal block feature vector is determined; Candidate event segments are determined based on spatial channels where the disturbance categories are continuous and identical in the spatial dimension, and the disturbance category is either mechanical or artificial. The effective event segments of the distributed optical fiber are obtained by filtering the candidate event segments based on their channel lengths.

[0013] Optionally, the step of performing vehicle diagonal texture filtering based on the effective event segment to obtain the warning result for the effective event segment includes: Identify the vehicle diagonal continuous texture in the two-dimensional waterfall image to obtain the vehicle texture influence segment of the distributed optical fiber; Based on the positional overlap relationship between the effective event segment and the vehicle texture influence segment, the overlapping segment of the effective event segment is determined; Overlapping segments of the valid event segments are removed or downweighted to obtain filtered valid event segments; The start and end channel numbers of the filtered valid event segments are converted into physical location information, and the warning results of the valid event segments are obtained based on the physical location information.

[0014] Secondly, the present invention also provides a distributed optical fiber vibration event identification and early warning system, comprising: The waterfall plot generation unit is used to acquire vibration data of distributed optical fibers in real time, and to arrange the vibration data into a matrix to obtain a two-dimensional waterfall plot of the vibration data. The feature extraction unit is used to perform dual-criterion denoising and feature extraction processing on the two-dimensional waterfall plot to obtain the physical statistical features of the two-dimensional waterfall plot. A vector construction unit is used to construct multiple spatiotemporal blocks based on the physical statistical characteristics, and obtain a spatiotemporal block feature vector corresponding to each spatiotemporal block; The prediction unit is used to learn a classification model through a preset incremental discrimination and make predictions based on the spatiotemporal block feature vectors to obtain the block-level recognition result of each spatiotemporal block feature vector. The correction unit is used to make decision corrections on each block-level recognition result to obtain the corrected block-level recognition result. The segment extraction unit is used to determine the valid event segments of the distributed optical fiber based on all the corrected block-level identification results. The warning generation unit is used to perform vehicle diagonal texture filtering based on the effective event segment to obtain the warning result of the effective event segment.

[0015] Thirdly, the present invention also provides an electronic device, comprising: a processor and a memory, wherein the memory is used to store a computer program; When the computer program is loaded by the processor, it causes the processor to execute the distributed optical fiber vibration event identification and early warning method as described above.

[0016] The distributed optical fiber vibration event identification and early warning method, system, and electronic device of the present invention first acquires vibration data in real time and arranges it into a two-dimensional waterfall plot, thereby preserving complete temporal and spatial information; then, the waterfall plot is subjected to dual-criteria denoising and feature extraction to remove invalid noise in low-quality channels and simultaneously extract statistical features with clear physical meaning; next, multiple spatiotemporal blocks are constructed based on the physical statistical features, and spatiotemporal block feature vectors are generated, thereby jointly representing the continuous expansion characteristics of the event in the temporal and spatial dimensions, avoiding the limitations of traditional single-point judgment; furthermore, a preset incremental discriminant learning classification model is used to predict the feature vector of each spatiotemporal block to obtain block-level recognition results, increasing... The quantitative discriminative learning classification model improves the stability of the classification boundary and its adaptability to boundary samples through inter-class difference guidance and incremental learning mechanisms. Based on this, decision corrections are made for each block-level identification result, further suppressing false alarms at the block level. Then, based on all corrected block-level identification results, the effective event segments of the distributed optical fiber are determined, resulting in continuous event intervals with engineering significance, eliminating discrete false alarms. Finally, vehicle diagonal texture filtering is applied to the effective event segments to remove false alarms caused by diagonal textures formed by vehicle traffic, ultimately outputting the warning result. This ensures the accuracy and reliability of the warning result, systematically improving the accuracy of identification and warning of vibration events in distributed optical fibers. Attached Figure Description

[0017] Figure 1 This is one of the flowcharts of the distributed optical fiber vibration event identification and early warning method according to an embodiment of the present invention; Figure 2 This is a second schematic flowchart of a distributed optical fiber vibration event identification and early warning method according to another embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a distributed optical fiber vibration event identification and early warning system according to another embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0019] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0020] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0021] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0022] Combination Figure 1 As shown in the figure, a distributed optical fiber vibration event identification and early warning method provided by an embodiment of the present invention includes: Vibration data of distributed optical fibers are acquired in real time, and the vibration data is arranged in a matrix to obtain a two-dimensional waterfall plot of the vibration data.

[0023] Specifically, the vibration data of the distributed optical fiber is acquired in real time and arranged into a two-dimensional waterfall plot. The continuously acquired raw time-domain signals are organized into a structured matrix according to the time dimension and the spatial channel dimension. Through matrix arrangement, the vibration response amplitude at each moment and each spatial location can be retained intuitively and completely, ensuring the synchronous alignment of time and space information and avoiding information loss due to data format disorder or dimensional misalignment. This allows subsequent denoising, feature extraction and event recognition to be carried out in a clear space-time coordinate system.

[0024] The two-dimensional waterfall plot is subjected to dual-criteria denoising and feature extraction to obtain the physical statistical features of the two-dimensional waterfall plot.

[0025] Specifically, firstly, dual-criteria denoising addresses the signal quality differences in different spatial channels along the optical fiber. By comprehensively considering the proportion of low-amplitude pulses and the number of high-amplitude pulses over a long period, it determines whether a channel is an invalid channel with a long-term low response and removes it from the data, thus avoiding contamination of subsequent feature distribution and classification decisions by low-quality channels. Based on this, feature extraction calculates statistical characteristics reflecting different levels such as vibration energy, impact intensity, pulse density, and waveform sharpness from the denoised two-dimensional waterfall plot. This transforms the original amplitude sequence into a more discriminative feature space, allowing the differences between artificial disturbances, mechanical disturbances, and background noise to be quantified.

[0026] Multiple spatiotemporal blocks are constructed based on the physical statistical characteristics, and a spatiotemporal block feature vector is obtained for each spatiotemporal block.

[0027] Specifically, since effective disturbances on distributed optical fibers often do not appear in isolation at a single sampling point or a single channel, but persist continuously over a period of time and exhibit similar vibration patterns in certain continuous regions within the distributed optical fiber, it is necessary to construct spatiotemporal blocks that can simultaneously cover the time window and spatial neighborhood. Then, the physical statistical features of each spatiotemporal block are flattened in a fixed order and spliced ​​into a high-dimensional vector to obtain the spatiotemporal block feature vector. This allows subsequent classification models to simultaneously see the continuity pattern of events in the time dimension and the expansion range in the spatial channel dimension, thereby changing the information limitations of traditional single-point or single-channel judgments and improving the ability to capture weak disturbances and boundary events.

[0028] By using a preset incremental discriminative learning classification model, prediction is performed based on the spatiotemporal block feature vectors to obtain the block-level recognition result for each spatiotemporal block feature vector.

[0029] Specifically, a pre-trained incremental discriminative learning classification model is used to classify the spatiotemporal block feature vectors corresponding to each spatiotemporal block, outputting whether the spatiotemporal block feature vector belongs to the background, artificial disturbance, or mechanical disturbance category. The incremental discriminative learning classification model differs from traditional random weight single-hidden-layer networks. Its hidden-layer weights are constructed guided by the direction of feature center differences between different categories, thus injecting prior information about class separability into the model and making the classification boundary more discriminative. Simultaneously, the incremental discriminative learning classification model also employs a closed-loop solution for output layer weights and a hard-sample incremental backfeeding training strategy, making it more adaptable and stable when facing boundary-ambiguous samples and weak event samples commonly encountered in engineering scenarios. By independently performing forward computation on the feature vector of each spatiotemporal block through the incremental learning classification model, the block-level recognition result is determined, realizing the mapping from feature space to event type.

[0030] Each block-level recognition result is then corrected to obtain the corrected block-level recognition result.

[0031] Specifically, the block-level recognition results directly output by the incremental discriminative learning classification model undergo secondary verification and adjustment to suppress unreliable judgments with low confidence or near boundaries, thereby reducing the negative impact of false alarms in block-level recognition results on the final warning results from the source. Through decision correction, the credibility of each block-level recognition result is effectively improved, providing higher-quality input for subsequent determination of valid event segments.

[0032] Based on all the corrected block-level identification results, the valid event segments of the distributed optical fiber are determined.

[0033] Specifically, the effective event segments of the distributed optical fiber are determined based on all the corrected block-level identification results. The results are then fused according to the spatial dimension, and continuous channel intervals of the same non-background category are searched in the spatial direction. Finally, continuous event segments with engineering significance are generated, thereby realizing the leap from micro-spatiotemporal block identification to macro-event localization and ensuring the stability and spatial continuity of the early warning results.

[0034] Based on the effective event segments, vehicle diagonal texture filtering is performed to obtain the warning results for the effective event segments.

[0035] Specifically, in real-world engineering environments, in areas frequently traversed by highways, construction roads, or vehicles, vehicle movement creates a distinct diagonal continuous texture in the waterfall diagram. Since this diagonal continuous texture may resemble mechanical or artificial disturbances in terms of energy and spatiotemporal continuity, it can easily lead to false alarms. In this embodiment, by filtering effective event segments using vehicle diagonal textures, the false alarm rate caused by vehicle interference is reduced, ensuring that the final output warning results have higher accuracy and on-site reliability, thus meeting the practical requirements of low false alarms in security linkage.

[0036] The distributed optical fiber vibration event identification and early warning method, system, and electronic device of the present invention first acquires vibration data in real time and arranges it into a two-dimensional waterfall plot, thereby preserving complete temporal and spatial information; then, the waterfall plot is subjected to dual-criteria denoising and feature extraction to remove invalid noise in low-quality channels and simultaneously extract statistical features with clear physical meaning; next, multiple spatiotemporal blocks are constructed based on the physical statistical features, and spatiotemporal block feature vectors are generated, thereby jointly representing the continuous expansion characteristics of the event in the temporal and spatial dimensions, avoiding the limitations of traditional single-point judgment; furthermore, a preset incremental discriminant learning classification model is used to predict the feature vector of each spatiotemporal block to obtain block-level recognition results, increasing... The quantitative discriminative learning classification model improves the stability of the classification boundary and its adaptability to boundary samples through inter-class difference guidance and incremental learning mechanisms. Based on this, decision corrections are made for each block-level identification result, further suppressing false alarms at the block level. Then, based on all corrected block-level identification results, the effective event segments of the distributed optical fiber are determined, resulting in continuous event intervals with engineering significance, eliminating discrete false alarms. Finally, vehicle diagonal texture filtering is applied to the effective event segments to remove false alarms caused by diagonal textures formed by vehicle traffic, ultimately outputting the warning result. This ensures the accuracy and reliability of the warning result, systematically improving the accuracy of identification and warning of vibration events in distributed optical fibers.

[0037] Optionally, the real-time acquisition of vibration data from the distributed optical fiber, and the matrix arrangement of the vibration data to obtain a two-dimensional waterfall plot of the vibration data, includes: Vibration data of the distributed optical fiber are acquired using a distributed optical fiber acoustic sensing device at a preset acquisition frequency. The distributed optical fiber is divided into multiple spatial channels according to a preset spatial resolution; Based on the vibration data, determine the vibration response amplitude of each of the spatial channels at each acquisition time; Based on the vibration response amplitude of each spatial channel at each acquisition time, a two-dimensional vibration matrix is ​​obtained by arranging the data matrix according to the time dimension and the spatial channel dimension. The two-dimensional vibration matrix is ​​used as the two-dimensional waterfall diagram.

[0038] Specifically, firstly, a distributed fiber optic acoustic sensing device continuously collects backscattered signals at various locations along the fiber optic cable at a preset sampling frequency, such as a sampling rate of several thousand to tens of thousands of hertz per second set according to monitoring requirements, thereby obtaining the original vibration time series. The entire distributed fiber optic cable is divided into multiple spatial channels according to a preset spatial resolution (e.g., one monitoring unit per meter, five meters, or ten meters). Each spatial channel corresponds to a continuous physical location on the fiber optic cable, serving as the basic granularity for subsequent vibration localization. Based on the collected vibration data, the vibration response amplitude of each spatial channel at each sampling moment is determined, i.e., the amplitude information belonging to the location range of this channel is extracted from the original signal. Then, based on the vibration response amplitude of each spatial channel at each sampling moment, a two-dimensional matrix is ​​arranged with the time dimension as the row direction (representing the chronological order of sampling moments) and the spatial channel dimension as the column direction (representing the arrangement order along the fiber optic distance), thereby obtaining a two-dimensional vibration matrix. Each element of the two-dimensional vibration matrix reflects the vibration intensity at a specific location at a specific moment. Finally, the two-dimensional vibration matrix is ​​directly used as a two-dimensional waterfall plot without the need for additional image rendering or format conversion, thus preserving the spatiotemporal structure of the original data.

[0039] The expression for the two-dimensional vibration matrix is: ; Where T represents the number of time sampling points corresponding to all acquisition times, and S represents the number of spatial channels. Let be the vibration response amplitude of the j-th spatial channel at the t-th acquisition time.

[0040] In a preferred embodiment of the present invention, a 10-kilometer-long oil and gas pipeline is selected as the monitoring object. Distributed fiber optic acoustic sensing equipment is used to continuously collect vibration data at a preset sampling frequency of 2000 Hz. The entire fiber optic cable is divided into 2000 spatial channels with a spatial resolution of 5 meters, each channel corresponding to a 5-meter-long section of the pipeline. For the collected raw vibration data, the vibration response amplitude of each spatial channel at each sampling time (one sampling point per millisecond) is extracted. For example, the amplitude at the 100th spatial channel and the 500th sampling time is 0.35 (relative unit). Then, 10,000 consecutive sampling times in the time dimension are used as rows of a matrix, and 2000 spatial channels are used as columns of the matrix, arranged in ascending order of time and channel number, to obtain a two-dimensional vibration matrix with 10,000 rows and 2000 columns.

[0041] In this embodiment of the invention, vibration data is acquired at a preset acquisition frequency and spatial channels are divided with a preset spatial resolution, thereby achieving flexible adaptation to the temporal and spatial accuracy of the monitoring scene. At the same time, by determining the vibration response amplitude of each spatial channel at each acquisition moment and arranging it into a two-dimensional matrix, the original signal is directly organized into a two-dimensional vibration matrix that retains the temporal and spatial correlation as a waterfall plot, thus avoiding information loss caused by data transformation.

[0042] Optionally, the step of performing dual-criterion denoising and feature extraction on the two-dimensional waterfall plot to obtain the physical statistical features of the two-dimensional waterfall plot includes: Based on the vibration response amplitude of each spatial channel in the two-dimensional waterfall diagram, determine the low amplitude ratio and high amplitude point count for each spatial channel; Based on the comparison results of the low amplitude ratio of the spatial channel with the preset low amplitude ratio threshold, and the comparison results of the high amplitude point count of the spatial channel with the preset high amplitude point quantity threshold, denoising processing is performed to obtain a denoised two-dimensional waterfall image. The denoised two-dimensional waterfall plot is extracted by sliding according to a time window of a preset size to obtain the short-time energy, high amplitude ratio, number of high amplitude continuous segments, root mean square of short-time energy, peak value of vibration response, and excess value of vibration response kurtosis for each spatial channel within each time window. The short-time energy, high amplitude ratio, number of continuous high amplitude segments, root mean square of short-time energy, peak value of vibration response, and excess value of vibration response kurtosis of all spatial channels within all time windows are used as the physical statistical characteristics of the two-dimensional waterfall plot.

[0043] Specifically, firstly, based on the vibration response amplitude of each spatial channel in the two-dimensional waterfall plot, the proportion of low amplitude points and the count of high amplitude points for each spatial channel are calculated. The proportion of low amplitude points is defined as the percentage of sampling points in that channel whose vibration amplitude is below the low amplitude threshold throughout the entire time period, calculated using the following formula: ; Where T represents the number of time sampling points corresponding to all acquisition times. Let be the vibration response amplitude of the j-th spatial channel at the t-th acquisition time. This is an indicator function (it takes the value 1 when the condition inside the parentheses is true, and 0 otherwise). The preset low amplitude threshold is (preferably the 10th percentile of the amplitude of all pixels in the entire waterfall image, with the lowest 10% amplitude being used as the background noise level). This represents the low amplitude ratio of the j-th spatial channel.

[0044] The high-amplitude point count is defined as the number of sampling points whose vibration amplitude exceeds the high-amplitude threshold throughout the entire time period of the channel. The calculation formula is as follows: ; Where T represents the number of time sampling points corresponding to all acquisition times. Let be the vibration response amplitude of the j-th spatial channel at the t-th acquisition time. This is an indicator function (it takes the value 1 when the condition inside the parentheses is true, and 0 otherwise). The high amplitude threshold is set (preferably the 90th percentile of the amplitude of all pixels in the entire waterfall image, with the highest 10% amplitude as the effective signal threshold). Count the high amplitude points of the j-th spatial channel.

[0045] Secondly, based on the low amplitude ratio of each spatial channel The comparison result with the preset low amplitude ratio threshold η (e.g., fixed at 0.8, or adaptively adjusted according to noise intensity), and the high amplitude point count of this spatial channel. The comparison result with the preset threshold M for the number of high-amplitude points is then subjected to noise reduction processing. When the condition is met... , At that time, the j-th spatial channel is set to zero. M is dynamically calculated based on 4% to 5% of the number of time sampling points T, i.e.: The longer the time period, the more high-value points are required to avoid misjudgment due to short-term noise.

[0046] Then, the denoised two-dimensional waterfall plot is extracted by sliding within a time window of a preset size. Within each time window, the physical statistical characteristics (short-time energy, proportion of high amplitude, number of continuous high amplitude segments, root mean square of short-time energy, peak value of vibration response, and excess value of vibration response kurtosis) are calculated for each spatial channel. Short-time energy reflects the overall intensity of vibration within a local time window, and its calculation formula is as follows: ; Where i is the index of the time window, u = i + P, u is the starting sampling point of the time window, P is the step size, and L is the length of the time window in the preset size. Let be the vibration response amplitude of the j-th spatial channel at the t-th acquisition time. Let L be the short-time energy of the j-th spatial channel within the i-th time window. In a preferred embodiment of the invention, L=16, P=1.

[0047] The proportion of high amplitude values ​​reflects the density of strong response points within the time window, and its calculation formula is as follows: ; in, The percentage of high amplitude values ​​in the j-th spatial channel within the i-th time window. This is the statistical threshold for high amplitude values.

[0048] The number of high-amplitude continuous segments reflects whether strong pulses appear in discrete clusters or repetitive segments, and is particularly sensitive to mechanical shock events. The calculation formula is as follows: ; in, This represents the number of consecutive high-amplitude segments in the j-th spatial channel within the i-th time window. To provide the high-amplitude mask for the j-th spatial channel at the t-th acquisition time, This is the high amplitude mask for the j-th spatial channel at the (t-1)-th acquisition time.

[0049] The root mean square of short-time energy reflects the level of the local amplitude envelope, and its calculation formula is as follows: ; in, The root mean square of the short-time energy of the j-th spatial channel within the i-th time window. Use a very small positive number to prevent numerical overflow. Let be the short-time energy of the j-th spatial channel within the i-th time window.

[0050] The peak vibration response reflects the extreme impact intensity within a local time window, and its calculation formula is as follows: ; in, Let be the vibration response amplitude of the j-th spatial channel at the t-th acquisition time, i be the index of the time window, u = i + P, u be the starting sampling point of the time window, and L be the length of the time window. Let be the peak value of the vibration response of the j-th spatial channel at the t-th acquisition time.

[0051] The vibration response kurtosis excess reflects the sharpness of the waveform and transient impact, playing an important role in distinguishing background noise from short-term impact disturbances. Its calculation formula is as follows: ; in, The vibration response kurtosis excess value of the j-th spatial channel within the i-th time window. Let be the average vibration amplitude of the j-th spatial channel within the i-th time window. Let be the vibration response amplitude of the j-th spatial channel at the t-th acquisition time. Let u be a very small positive number to prevent numerical overflow, where u is the starting sampling point of the time window, and L is the length of the time window.

[0052] Finally, the six types of feature values ​​calculated for all spatial channels within all time windows are used together as the physical statistical features of the two-dimensional waterfall plot, forming a feature tensor for subsequent spatiotemporal block construction.

[0053] In this embodiment of the invention, by employing a combined criterion of low amplitude ratio and high amplitude point count for channel-level denoising, pseudo-silent columns or low-noise columns with long-term low response and lack of effective pulses can be identified and eliminated. This avoids the blurring of event boundaries caused by traditional smoothing filters and suppresses the contamination of feature distribution by low-quality channels. By extracting six types of physical statistical features, including short-time energy, high amplitude ratio, number of continuous high amplitude segments, root mean square of short-time energy, peak value of vibration response, and excess value of vibration response kurtosis, the physical properties of distributed optical fiber vibration signals are comprehensively characterized from multiple dimensions such as vibration intensity, pulse density, impact structure, and kurtosis. This enhances the separability between artificial disturbances, mechanical disturbances, and background noise, thereby providing a highly discriminative feature basis for subsequent event recognition and significantly improving recognition accuracy and robustness.

[0054] Optionally, the step of constructing multiple spatiotemporal blocks based on the physical statistical characteristics to obtain a spatiotemporal block feature vector corresponding to each spatiotemporal block includes: The clipping window of the spatiotemporal block is determined based on the preset feature frame domain in the time dimension and the preset neighborhood range in the spatial channel dimension. Based on the clipping window, multiple spatiotemporal blocks are extracted from the physical statistical features; The spatial channels in the neighborhood of the spatiotemporal block are spliced ​​together in a preset order according to the short-time energy, high amplitude ratio, number of high amplitude continuous segments, root mean square of short-time energy, peak value of vibration response, and excess value of vibration response kurtosis in all time windows of the feature frame domain corresponding to the spatiotemporal block, to obtain the spatiotemporal block feature vector corresponding to the spatiotemporal block.

[0055] Specifically, firstly, based on the preset feature frame domain in the time dimension and the preset neighborhood range in the spatial channel dimension, the clipping window for each spatiotemporal block is determined. The feature frame domain refers to a fixed number of consecutively selected feature frames along the time direction on the feature frame coordinate axis formed after feature extraction. Each feature frame corresponds to a sliding time window (e.g., window length 16, step size 1). To determine the time width of the feature frame domain, a half-window width is introduced. ; When the step size P=1, the half-window width =250, total frame window width =2 =500, meaning each spatiotemporal block covers 500 consecutive feature frames in the time direction, corresponding to the time span of approximately 500 original sampling points in the original waterfall plot; the neighborhood range refers to the set of consecutive spatial channels formed by expanding to the left and right sides of the currently interested spatial channel on the spatial channel coordinate axis. The preferred neighborhood range is 6 spatial channels (for example, 3 on each side of the central channel c, or specifically [c-3, c+2]). The clipping window thus determined is a spatiotemporal frame with a fixed size (e.g., 500 feature frames × 6 spatial channels). Next, according to this clipping window, multiple spatiotemporal blocks are extracted from the extracted physical statistical feature tensors, that is, for each central feature frame and each central spatial channel, a local sub-tensor of a corresponding size is clipped. In the preferred embodiment of the present invention, each sub-tensor contains short-time energy E. block High amplitude accounts for R block Number of high-amplitude continuous segments N block Short-time energy root mean square (RMS) block Vibration response peak P block and vibration response kurtosis excess value K block The feature sub-blocks are defined as follows: These six types of feature sub-blocks represent all values ​​within the spatiotemporal range, with each sub-block covering both the feature frame domain and its neighborhood. Then, for each spatiotemporal block, the six types of feature values ​​within all time windows of the corresponding feature frame domain in its corresponding neighborhood are concatenated in a preset order. First, each type of feature sub-block (e.g., 500×6 in size) is flattened into a vector using either row-major or column-major order, denoted as vec(E...). block ),vec(R block ), vec(N block ), vec (RMS) block ),vec(P block ), vec(K block Each flattened vector has a dimension of 500 × 6 = 3000. These six vectors are then concatenated in a fixed order (e.g., first short-time energy, then the proportion of high amplitude, then the number of consecutive high amplitude segments, then the root mean square of short-time energy, then the peak value of vibrational response, and finally the excess value of vibrational response kurtosis) to form the final spatiotemporal block feature vector F. F=[vec(E block ),vec(R block ),vec(N block ),vec(RMS block ),vec(P block ),vec(K block )).

[0056] In the preferred embodiment with step size P=1, the time window corresponds to 500 feature frame positions, the spatial width is 6, the feature types are 6, and the input dimension of the feature vector is D=500×6×6=18000.

[0057] In this embodiment of the invention, by determining the spatiotemporal block clipping window according to the preset feature frame domain and neighborhood range, each spatiotemporal block can cover a sufficiently long time span and a sufficiently wide spatial neighborhood, thereby completely capturing the temporal persistence and spatial extensibility of distributed optical fiber vibration events. By flattening and splicing the six types of physical statistical feature sub-blocks into a high-dimensional vector in a fixed order, the multi-dimensional information such as the intensity of the event, pulse density, impact structure, and sharpness is integrated into a unified feature representation, avoiding the information fragmentation of traditional single-point or single-channel features. This spatiotemporal block feature vector serves as a recognition unit, enabling the classification model to simultaneously perceive the global pattern within the local spatiotemporal window, enhancing the ability to distinguish between weak disturbances, intermittent disturbances, and complex background interference.

[0058] Optionally, the step of predicting based on the spatiotemporal block feature vectors using a preset incremental discriminative learning classification model to obtain the block-level recognition result for each spatiotemporal block feature vector includes: The spatiotemporal block feature vector is input into the incremental discriminant learning classification model; The probability value of the spatiotemporal block feature vector under a preset category is obtained by calculating the probability value of the spatiotemporal block feature vector based on the hidden mapping layer and output layer of the incremental discriminative learning classification model, according to the spatiotemporal block feature vector. The preset category includes background class, artificial disturbance class and mechanical disturbance class. By comparing the probability values ​​corresponding to the background class, the artificial disturbance class, and the mechanical disturbance class, the disturbance category corresponding to the spatiotemporal block feature vector is determined, and the disturbance category is used as the block-level recognition result of the spatiotemporal block feature vector.

[0059] Specifically, firstly, the spatiotemporal block feature vector corresponding to each spatiotemporal block is input into a pre-trained incremental discriminative learning classification model. The incremental discriminative learning classification model adopts a single hidden layer network structure, and its hidden layer weights are not completely randomly initialized, but are composed of a mixture of inter-class difference vector-guided part and random generalization weight part.

[0060] Let the training set sample matrix be... The dimension is N×D, where N is the total number of training samples and D is the feature dimension of each sample. The mean vector μ and standard deviation vector σ of the training set sample matrix are calculated column-wise, and the samples are standardized, i.e.: ; in, This is the standardized training sample matrix, with each feature column having a mean of 0 and a standard deviation of 1, used to eliminate the impact of differences in the units of different features on model training; It is a very small positive number used to prevent numerical overflow caused by a denominator of zero.

[0061] Calculate the category centers for the background class (label 0), the artificial disturbance class (label 1), and the mechanical disturbance class (label 2) respectively: ; in, As the category center of class c, This represents the number of samples in class c, where c is the class index, representing the background class (label 0), the artificial disturbance class (label 1), and the mechanical disturbance class (label 2), respectively. Let g be the standardized feature vector of the g-th training sample, with dimension 1×D. Let g be the standardized spatiotemporal block feature vector of the g-th sample.

[0062] For any two classes (a, b), construct the normalized inter-class difference vector as follows: ; in, Let be the normalized inter-class difference vector pointing from class b to class a. Both a and b take values ​​of 0, 1, and 2, and a ≠ b. When a = 1 and b = 0, it represents the direction of difference between the artificial perturbation class and the background class. Let be the feature center vector of the a-th sample. Let be the feature center vector of the b-th sample. This is a very small positive number used to prevent numerical overflow caused by a zero denominator. The normalized inter-class difference vector represents the main separating direction of different classes in the high-dimensional feature space and is used to guide the initialization of hidden layer weights.

[0063] In this embodiment, the number of hidden layer nodes is set to H. n Hidden layer weight matrix It is divided into two parts, namely the difference-guided weight submatrix. and random generalization weight submatrix It is pieced together, that is The difference-guided weights are composed of inter-class difference vectors, the random generalization weights are composed of random vectors, and the hidden layer bias vector b is initialized with a small amount of randomness. Next, the spatiotemporal block feature vectors of the input are calculated through the model's hidden mapping layer and output layer. Specifically, in the hidden mapping layer, the input spatiotemporal block feature vectors are calculated... Perform a linear transformation and apply an activation function to obtain the hidden layer output matrix: ; in, for The function, where b is the hidden layer bias vector. Let Y be the hidden layer weight matrix and H be the hidden layer output matrix. In the output layer, to alleviate the crowding out of minority classes caused by an excessive number of background class samples, class weights are assigned to the one-hot label encoding matrix. Let Y be the original one-hot label matrix, then the weighted label matrix is: ; in, For a weighted label matrix, This is the original one-hot encoding matrix for the labels. Let the class weights be , where For background class weights, For artificially perturbed weights, Weights are for mechanical disturbances. It is a diagonal matrix, with diagonal elements representing class weights. .

[0064] Then, the output layer weights are solved using a closed-form algorithm. When the sample size is large, a fast response matrix decomposition is performed on the hidden layer output matrix H, i.e.: Where Q is an orthogonal matrix with dimensions N×H n H is the hidden layer output matrix with dimensions N×H. n R is an upper triangular matrix with dimension H. n ×H n The output layer weights are solved using QR decomposition. When the sample size is small, the generalized inverse can be used directly to solve the problem. ; in, The output layer weight matrix has dimension H. n ×3; It is the inverse of the upper triangular matrix R. It is the transpose of an orthogonal matrix. The generalized inverse of the hidden layer output matrix. This is a weighted label matrix with dimensions N×3.

[0065] The output layer weight matrix β is obtained through the closed-form solution described above. Finally, three probability values ​​are calculated for the spatiotemporal block feature vector to belong to the background class, the artificial disturbance class, and the mechanical disturbance class, respectively. The three probability values ​​are then compared, and the preset category with the highest probability value is taken as the disturbance category corresponding to the spatiotemporal block feature vector. This disturbance category is then output as the block-level recognition result.

[0066] In this embodiment of the invention, an incremental discriminative learning classification model that uses inter-class difference vectors to guide the initialization of hidden layers is adopted, enabling the model to obtain directional priors for class separation during the training phase. Compared with the traditional single-hidden-layer model with completely random initialization, it has better classification boundary stability. By using hidden layer mapping and closed output layer calculation, the backpropagation iteration of deep networks is avoided, achieving efficient forward inference. By outputting three class probabilities and taking the maximum value to determine the block-level recognition result, a preliminary class judgment can be quickly made for each local spatiotemporal window, thereby improving the overall recognition accuracy while ensuring the real-time performance of inference.

[0067] Optionally, the step of making decision corrections for each block-level recognition result to obtain corrected block-level recognition results includes: When the perturbation category of the spatiotemporal block feature vector is the artificial perturbation class, if the probability value of the spatiotemporal block feature vector in the artificial perturbation class is less than the preset minimum artificial confidence level, then the perturbation category of the spatiotemporal block feature vector is corrected to the background class, and the corrected block-level recognition result is obtained. When the perturbation category of the spatiotemporal block feature vector is the mechanical perturbation category, if the probability value of the spatiotemporal block feature vector in the mechanical perturbation category is less than the preset minimum confidence threshold for machinery, then the perturbation category of the spatiotemporal block feature vector is corrected to the background category, and the corrected block-level recognition result is obtained. When the probability value of the spatiotemporal block feature vector in the mechanical disturbance class is greater than or equal to a preset mechanical forced reclassification threshold, and the difference between the probability value of the spatiotemporal block feature vector in the mechanical disturbance class and the probability value of the spatiotemporal block feature vector in the background class is greater than or equal to a preset mechanical relative advantage margin, and the probability value of the spatiotemporal block feature vector in the mechanical disturbance class is greater than or equal to the probability value of the spatiotemporal block feature vector in the artificial disturbance class, then the disturbance category of the spatiotemporal block feature vector is corrected to the mechanical disturbance class, and the corrected block-level recognition result is obtained.

[0068] Specifically, firstly, based on the three probability values ​​output by the classification model (the probability value of the background class)... Probability values ​​of artificial disturbances Probability values ​​of mechanical disturbances The initial perturbation category of the current spatiotemporal block is determined. Then, three preset multi-threshold rules are used to correct the initial category. The first rule targets artificial perturbation: when the initial category is artificial perturbation, it is determined that... Is it less than the preset minimum confidence level? (The preferred value is 0.6). If it is less than 0.6, the disturbance category of the spatiotemporal block is reverted from the artificial disturbance category to the background category, resulting in a corrected block-level recognition result. The second rule applies to the mechanical disturbance category: when the initial category is the mechanical disturbance category, it is determined that... Is it less than the preset minimum confidence threshold for machinery? (The preferred value is 0.65). If it is less than 0.65, the perturbation category of the spatiotemporal block will be reverted from the mechanical perturbation category to the background category. The third rule is the mechanical forced correction rule: when the following three conditions are met simultaneously, regardless of the perturbation category of the spatiotemporal block feature vector, the perturbation category will be forcibly corrected to the mechanical perturbation category. Condition 1: Mechanical perturbation category probability value Greater than or equal to the preset mechanical forced judgment threshold (The preferred value is 0.8); Condition 2, the difference between the probability value of mechanical disturbance and the probability value of background disturbance. Greater than or equal to the preset relative advantage margin of machinery (The preferred value is 0.3); Condition 3, the probability value of mechanical disturbance. The probability value is greater than or equal to that of the artificial disturbance class. By applying the above three rules, the block-level identification results of each spatiotemporal block are re-evaluated and corrected, ultimately outputting a more reliable corrected block-level identification result.

[0069] In this embodiment of the invention, by introducing a manual minimum confidence threshold and a mechanical minimum confidence threshold, the low-confidence initial category output by the model is backtracked, effectively eliminating boundary samples where the model's judgment confidence is insufficient, thus reducing block-level false alarms from the source. By introducing a mechanical forced reclassification rule (including a mechanical forced reclassification threshold, a mechanical relative advantage margin, and a comparison with the manual category), even if the initial category is not a mechanical perturbation, as long as the probability of mechanical perturbation is significantly dominant and reaches a high confidence level, it can be forcibly corrected to the mechanical category, thereby improving the detection rate of mechanical perturbations. While maintaining a high recall rate, the false alarm rate is significantly reduced, making the block-level identification results of each spatiotemporal block more reliable.

[0070] Optionally, determining the valid event segments of the distributed optical fiber based on all the corrected block-level identification results includes: Based on the corrected block-level identification results of all the spatiotemporal block feature vectors, the disturbance category of the spatial channel contained in the spatiotemporal block corresponding to each spatiotemporal block feature vector is determined; Candidate event segments are determined based on spatial channels where the disturbance categories are continuous and identical in the spatial dimension, and the disturbance category is either mechanical or artificial. The effective event segments of the distributed optical fiber are obtained by filtering the candidate event segments based on their channel lengths.

[0071] Specifically, since each spatial channel corresponds to multiple spatiotemporal blocks on different time center frames, and each spatiotemporal block obtains a corrected block-level identification result after the aforementioned decision correction (i.e., the disturbance category to which the spatiotemporal block belongs, including background, artificial disturbance, or mechanical disturbance), it is necessary to summarize all corrected block-level identification results according to their respective spatial channels. Specifically, for each spatial channel, the corrected disturbance categories of all its corresponding spatiotemporal blocks are collected, and the comprehensive category of the spatial channel in the time dimension is determined through column-level voting aggregation operations (e.g., counting the frequency of occurrence of each category or calculating the average probability), i.e., the column-level label of each spatial channel. This column-level label reflects the disturbance category ultimately determined for the spatial channel over a period of time. Next, after obtaining the column-level labels of all spatial channels, the label sequence is scanned along the spatial dimension (i.e., the fiber length direction) to find spatial channel intervals where the disturbance category is consecutively the same and that category is mechanical disturbance or artificial disturbance (not background). These consecutive intervals are determined as candidate event segments. Each candidate event segment includes a start channel number, an end channel number, and the number of consecutive channels. Finally, for each candidate event segment, its continuous channel length is compared with the minimum continuous length threshold corresponding to artificial disturbance or mechanical disturbance, respectively. If the channel length of the candidate event segment is greater than or equal to the minimum continuous length threshold (for example, the minimum continuous length threshold for artificial disturbance can be set to 2 channels, and the minimum continuous length threshold for mechanical disturbance can be set to 3 channels), it is determined as a valid event segment and used as a valid event segment of the distributed optical fiber; if the length is insufficient, it is regarded as isolated noise and discarded.

[0072] In a preferred embodiment of the present invention, the disturbance category of a spatial channel can be determined by column-level voting. Specifically, for a given spatial channel, let the block-level labels of all its corresponding spatiotemporal blocks be... The expression for calculating the number of votes is: ; Where k∈{0,1,2} represents the background class, the artificial disturbance class, and the mechanical disturbance class, respectively, and the probability mean of each class is calculated. For spacetime blocks, the block-level tag. Let 'b' be the index of the spatiotemporal block, and let 'b' be the number of votes for each channel in each category. Further calculations are made to determine the proportions of mechanical, manual, and background votes. Then, a comprehensive decision is made by combining these proportions with the average probability of each category and the proportion of strong blocks. Specifically, if both the proportion of mechanical votes and the proportion of strong mechanical blocks are high, and the average probability of the mechanical disturbance category is higher than the average probability of the background category, the spatial channel is classified as a mechanical disturbance category. If both the proportion of manual votes and the proportion of strong manual blocks are high, and the average probability of the manual disturbance category meets a preset condition, the spatial channel is classified as a manual disturbance category. If the proportion of background votes is the highest or the average probability of the background category is not lower than other categories, the spatial channel is classified as a background category. If none of the above rules are sufficient, a weighted secondary judgment can be made by combining the number of votes and the probability value. Through this column-level voting aggregation, each spatial channel obtains a stable and reliable column-level label, which is the final disturbance category of the channel in the time dimension.

[0073] In this embodiment of the invention, by performing column-level voting aggregation on the corrected perturbation categories of multiple spatiotemporal blocks in each spatial channel, redundant information in the time dimension is fully utilized to suppress accidental misjudgments of individual spatiotemporal blocks, thereby obtaining stable and reliable column-level labels for each spatial channel. By searching for continuous channel intervals with the same non-background category along the spatial dimension and filtering based on the length of continuous channels, event segments that are truly caused by artificial or mechanical perturbations and extend continuously in space are retained, while isolated single-channel false alarms are effectively eliminated. This integrates discrete block-level identification results into spatially continuous event segments, enhancing the spatial continuity and engineering reliability of the warning results, and solving the problem of frequent discrete false alarms in existing technologies.

[0074] Optionally, the step of performing vehicle diagonal texture filtering based on the effective event segment to obtain the warning result for the effective event segment includes: Identify the vehicle diagonal continuous texture in the two-dimensional waterfall image to obtain the vehicle texture influence segment of the distributed optical fiber; Based on the positional overlap relationship between the effective event segment and the vehicle texture influence segment, the overlapping segment of the effective event segment is determined; Overlapping segments of the valid event segments are removed or downweighted to obtain filtered valid event segments; The start and end channel numbers of the filtered valid event segments are converted into physical location information, and the warning results of the valid event segments are obtained based on the physical location information.

[0075] Specifically, the absolute average energy of the two-dimensional data is calculated column by column to remove columns with strong vertical noise; two-dimensional median filtering is performed on the processing results to eliminate isolated impulse noise; grayscale normalization is performed on the filtered results to map them to a uniform grayscale range; diagonal textures are enhanced using convolution kernels at 45 degrees and 135 degrees; the enhanced results are binarized; straight line segments are extracted using probabilistic Hough linear transform; diagonal lines belonging to vehicle traffic are selected based on the tilt angle (close to 45° or 135°), horizontal length (projection in the spatial channel direction), and vertical length (projection in the time direction); each selected diagonal line is extended into a continuous segment in the spatial channel direction (it can be extended outward by a certain buffer distance), and spatially overlapping or connected segments are merged to finally obtain the vehicle texture influence segment of the distributed optical fiber. Secondly, the identified valid event segments (each valid event segment includes a start channel number, an end channel number, and an event type) are compared with the vehicle texture influence segments in terms of location. Based on the overlap relationship between the two in the spatial channel dimension, the segments in the valid event segments that overlap with the vehicle texture influence segments (i.e., overlapping segments) are determined. Then, the overlapping segments in the valid event segments are either eliminated or downweighted: if the overlapping segment covers the entire valid event segment, the event segment is completely eliminated; if there is only partial overlap, the overlapping part can be eliminated while the non-overlapping part is retained, or the confidence of the entire event segment can be downweighted. After this processing, the filtered valid event segments are obtained. Finally, the start channel number and end channel number of the filtered valid event segments are converted into physical location information. The specific conversion method is as follows: assuming the total monitoring length of the distributed optical fiber is... The number of spatial channels is S, and the formula for calculating the physical distance d corresponding to a single channel is: ; For the starting channel number s and the ending channel number e, calculate the starting physical position and the ending physical position respectively: ; ; ; in, The starting physical location, To end the physical location, Using the starting physical location, ending physical location, and center location as the center position, the system generates early warning results for valid event segments based on the aforementioned physical location information, event type (human disturbance or mechanical disturbance), and occurrence time. These results are then output to the security linkage system.

[0076] In this embodiment of the invention, by identifying the vehicle's diagonal continuous texture in the two-dimensional waterfall diagram and generating vehicle texture influence segments, typical diagonal texture patterns formed by vehicle passage can be detected in a targeted manner, effectively distinguishing vehicle background interference from truly dangerous manual construction or mechanical excavation events; by judging the position overlap with the effective event segments and removing or downweighting the overlapping segments, the false alarm rate caused by vehicle passage is significantly reduced, avoiding invalid responses to false vehicle alarms at the engineering site; by converting the channel number of the filtered effective event segments into actual physical location information, the warning results can intuitively provide the accurate geographical location of the event (e.g., pipeline mileage or coordinates), meeting the engineering requirements of the security linkage system for alarm location accuracy and warning reliability, thereby ensuring the on-site availability of the final warning results.

[0077] In summary, combining Figure 2As shown, in a preferred embodiment of the present invention, raw vibration data along the line is first collected using distributed fiber optic acoustic sensing devices, and the data is acquired and decoded to obtain a two-dimensional waterfall plot. Next, channel-level dual-criteria denoising is performed, that is, the low-amplitude ratio and high-amplitude point count are calculated for each spatial channel, and channels that simultaneously satisfy the condition of a low-amplitude ratio higher than a threshold and a high-amplitude point count lower than a threshold are set to zero, completing channel-level dual-column joint decoding. Then, a sliding time window (window length 16, step size 1) is used to extract six types of physical statistical features for each spatial channel within each time window: short-time energy, high-amplitude ratio, number of consecutive high-amplitude segments, root mean square, peak value, and kurtosis excess value, forming a feature tensor. Based on this, a 500×6 spatiotemporal block is constructed by covering 500 feature frames in the temporal direction and 6 adjacent channels in the spatial direction. The six types of feature sub-blocks within each spatiotemporal block are flattened and stitched together to obtain an 18000-dimensional spatiotemporal block feature vector. During the training phase, samples filtered based on label consistency are used. Hidden layer weights are initialized using a hybrid approach guided by inter-class difference vectors. A closed-loop solution is employed for the output layer, followed by two rounds of incremental backfeeding with hard samples. This completes the training of a spatiotemporal difference-guided single-layer hidden layer classification model (an incremental discriminative learning classification model). After training, an optimal decision threshold search is performed to determine the decision boundary that optimizes crack recognition performance on the validation set. In the online inference phase, the feature vectors of the spatiotemporal blocks to be predicted are input into the trained model to obtain probabilities for three classes: background, human, and mechanical. Subsequently, block-level multi-threshold correction is performed: if the probability of the human class is lower than the preset minimum confidence level for human, it reverts to background; if the probability of the mechanical class is lower than the preset minimum confidence level for mechanical, it reverts to background; if the probability of the mechanical class satisfies high confidence, relative advantage over background, and relative advantage over human, it is forcibly classified as mechanical. For each spatial channel, the corrected results of all its spatiotemporal blocks are statistically analyzed, and the column-level label for that channel is determined through column-level voting (calculating the number of votes for each class and the average probability). Next, the column-level label sequence is scanned along the spatial dimension to extract consecutive intervals of the same non-background category. Intervals whose length reaches a preset minimum consecutive length threshold (≥2 channels for manual, ≥3 channels for mechanical) are determined as valid event segments (continuous segment determination). Then, vehicle trajectory texture filtering is performed: median filtering, grayscale normalization, directional convolution enhancement, and probabilistic Hough line detection are applied to the waterfall plot to identify 45° or 135° diagonal continuous textures and expand them into vehicle texture influence segments. Valid event segments falling into these segments are removed or downweighted. Finally, the start and end channel numbers of the filtered valid event segments are converted into physical location information, and a warning result containing event type, location, and time is output. Simultaneously, throughout the entire process, training management and model solidification are responsible for saving standardized parameters, thresholds, and model weights for repeated use in online inference.

[0078] Combination Figure 3 As shown, another embodiment of the present invention provides a distributed optical fiber vibration event identification and early warning system, comprising: The waterfall plot generation unit is used to acquire vibration data of distributed optical fibers in real time, and to arrange the vibration data into a matrix to obtain a two-dimensional waterfall plot of the vibration data. The feature extraction unit is used to perform dual-criterion denoising and feature extraction processing on the two-dimensional waterfall plot to obtain the physical statistical features of the two-dimensional waterfall plot. A vector construction unit is used to construct multiple spatiotemporal blocks based on the physical statistical characteristics, and obtain a spatiotemporal block feature vector corresponding to each spatiotemporal block; The prediction unit is used to learn a classification model through a preset incremental discrimination and make predictions based on the spatiotemporal block feature vectors to obtain the block-level recognition result of each spatiotemporal block feature vector. The correction unit is used to make decision corrections on each block-level recognition result to obtain the corrected block-level recognition result. The segment extraction unit is used to determine the valid event segments of the distributed optical fiber based on all the corrected block-level identification results. The warning generation unit is used to perform vehicle diagonal texture filtering based on the effective event segment to obtain the warning result of the effective event segment.

[0079] The distributed optical fiber vibration event identification and early warning system of the present invention has the same advantages over the prior art as the aforementioned distributed optical fiber vibration event identification and early warning method, and will not be repeated here.

[0080] Another embodiment of the present invention provides an electronic device, comprising: a processor and a memory, wherein the memory is used to store a computer program; When the computer program is loaded by the processor, it causes the processor to execute the distributed optical fiber vibration event identification and early warning method as described above.

[0081] The electronic device of the present invention has the same advantages over the prior art as the aforementioned distributed optical fiber vibration event identification and early warning method, and will not be repeated here.

[0082] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A distributed optical fiber vibration event identification and early warning method, characterized in that, include: Vibration data of distributed optical fibers are acquired in real time, and the vibration data is arranged in a matrix to obtain a two-dimensional waterfall plot of the vibration data. The two-dimensional waterfall plot is subjected to dual-criteria denoising and feature extraction to obtain the physical statistical features of the two-dimensional waterfall plot; Multiple spatiotemporal blocks are constructed based on the physical statistical characteristics, and a spatiotemporal block feature vector corresponding to each spatiotemporal block is obtained; By using a preset incremental discriminant learning classification model, prediction is performed based on the spatiotemporal block feature vectors to obtain block-level recognition results for each spatiotemporal block feature vector. Specifically, this includes: inputting the spatiotemporal block feature vectors into the incremental discriminant learning classification model; calculating the probability value of the spatiotemporal block feature vectors under preset categories through the hidden mapping layer and output layer of the incremental discriminant learning classification model, wherein the preset categories include background, artificial disturbance, and mechanical disturbance; comparing the probability values ​​corresponding to the background, artificial disturbance, and mechanical disturbance categories respectively to determine the disturbance category corresponding to the spatiotemporal block feature vectors, and using the disturbance category as the block-level recognition result of the spatiotemporal block feature vectors. For each block-level recognition result, a decision correction is performed to obtain a corrected block-level recognition result. Specifically, this includes: when the perturbation category of the spatiotemporal block feature vector is the artificial perturbation class, if the probability value of the spatiotemporal block feature vector in the artificial perturbation class is less than a preset minimum artificial confidence level, then the perturbation category of the spatiotemporal block feature vector is corrected to the background class, thus obtaining the corrected block-level recognition result; when the perturbation category of the spatiotemporal block feature vector is the mechanical perturbation class, if the probability value of the spatiotemporal block feature vector in the mechanical perturbation class is less than a preset minimum mechanical confidence level threshold, then the perturbation category of the spatiotemporal block feature vector is corrected to the background class. For the background class, the corrected block-level recognition result is obtained; when the probability value of the spatiotemporal block feature vector in the mechanical disturbance class is greater than or equal to the preset mechanical forced re-judgment threshold, and the difference between the probability value of the spatiotemporal block feature vector in the mechanical disturbance class and the probability value of the spatiotemporal block feature vector in the background class is greater than or equal to the preset mechanical relative advantage margin, and the probability value of the spatiotemporal block feature vector in the mechanical disturbance class is greater than or equal to the probability value of the spatiotemporal block feature vector in the artificial disturbance class, then the disturbance category of the spatiotemporal block feature vector is corrected to the mechanical disturbance class, and the corrected block-level recognition result is obtained; Based on all the corrected block-level identification results, the valid event segments of the distributed optical fiber are determined; Based on the effective event segments, vehicle diagonal texture filtering is performed to obtain the warning results for the effective event segments.

2. The distributed optical fiber vibration event identification and early warning method according to claim 1, characterized in that, The process of acquiring vibration data from distributed optical fibers in real time and arranging the vibration data in a matrix to obtain a two-dimensional waterfall plot of the vibration data includes: Vibration data of the distributed optical fiber are acquired using a distributed optical fiber acoustic sensing device at a preset acquisition frequency. The distributed optical fiber is divided into multiple spatial channels according to a preset spatial resolution; Based on the vibration data, determine the vibration response amplitude of each of the spatial channels at each acquisition time; Based on the vibration response amplitude of each spatial channel at each acquisition time, a two-dimensional vibration matrix is ​​obtained by arranging the data matrix according to the time dimension and the spatial channel dimension. The two-dimensional vibration matrix is ​​used as the two-dimensional waterfall diagram.

3. The distributed optical fiber vibration event identification and early warning method according to claim 2, characterized in that, The process of performing dual-criteria denoising and feature extraction on the two-dimensional waterfall plot to obtain the physical statistical features of the two-dimensional waterfall plot includes: Based on the vibration response amplitude of each spatial channel in the two-dimensional waterfall diagram, determine the low amplitude ratio and high amplitude point count for each spatial channel; Based on the comparison results of the low amplitude ratio of the spatial channel with the preset low amplitude ratio threshold, and the comparison results of the high amplitude point count of the spatial channel with the preset high amplitude point quantity threshold, denoising processing is performed to obtain a denoised two-dimensional waterfall image. The denoised two-dimensional waterfall plot is extracted by sliding according to a time window of a preset size to obtain the short-time energy, high amplitude ratio, number of high amplitude continuous segments, root mean square of short-time energy, peak value of vibration response, and excess value of vibration response kurtosis for each spatial channel within each time window. The short-time energy, high amplitude ratio, number of continuous high amplitude segments, root mean square of short-time energy, peak value of vibration response, and excess value of vibration response kurtosis of all spatial channels within all time windows are used as the physical statistical characteristics of the two-dimensional waterfall plot.

4. The distributed optical fiber vibration event identification and early warning method according to claim 3, characterized in that, The step of constructing multiple spatiotemporal blocks based on the physical statistical characteristics, and obtaining a spatiotemporal block feature vector corresponding to each spatiotemporal block, includes: The clipping window of the spatiotemporal block is determined based on the preset feature frame domain in the time dimension and the preset neighborhood range in the spatial channel dimension. Based on the clipping window, multiple spatiotemporal blocks are extracted from the physical statistical features; The spatial channels in the neighborhood of the spatiotemporal block are spliced ​​together in a preset order according to the short-time energy, high amplitude ratio, number of high amplitude continuous segments, root mean square of short-time energy, peak value of vibration response, and excess value of vibration response kurtosis in all time windows of the feature frame domain corresponding to the spatiotemporal block, to obtain the spatiotemporal block feature vector corresponding to the spatiotemporal block.

5. The distributed optical fiber vibration event identification and early warning method according to claim 1, characterized in that, The determination of valid event segments of the distributed optical fiber based on all the corrected block-level identification results includes: Based on the corrected block-level identification results of all the spatiotemporal block feature vectors, the disturbance category of the spatial channel contained in the spatiotemporal block corresponding to each spatiotemporal block feature vector is determined; Candidate event segments are determined based on spatial channels where the disturbance categories are continuous and identical in the spatial dimension, and the disturbance category is either mechanical or artificial. The effective event segments of the distributed optical fiber are obtained by filtering the candidate event segments based on their channel lengths.

6. The distributed optical fiber vibration event identification and early warning method according to claim 5, characterized in that, The step of performing vehicle diagonal texture filtering based on the effective event segments to obtain the warning result for the effective event segments includes: Identify the vehicle diagonal continuous texture in the two-dimensional waterfall image to obtain the vehicle texture influence segment of the distributed optical fiber; Based on the positional overlap relationship between the effective event segment and the vehicle texture influence segment, the overlapping segment of the effective event segment is determined; Overlapping segments of the valid event segments are removed or downweighted to obtain filtered valid event segments; The start and end channel numbers of the filtered valid event segments are converted into physical location information, and the warning results of the valid event segments are obtained based on the physical location information.

7. A distributed optical fiber vibration event identification and early warning system, characterized in that, include: The waterfall plot generation unit is used to acquire vibration data of distributed optical fibers in real time, and to arrange the vibration data into a matrix to obtain a two-dimensional waterfall plot of the vibration data. The feature extraction unit is used to perform dual-criterion denoising and feature extraction processing on the two-dimensional waterfall plot to obtain the physical statistical features of the two-dimensional waterfall plot. A vector construction unit is used to construct multiple spatiotemporal blocks based on the physical statistical characteristics, and obtain a spatiotemporal block feature vector corresponding to each spatiotemporal block; The prediction unit is used to predict the spatiotemporal block feature vectors based on the preset incremental discriminant learning classification model to obtain the block-level recognition result of each spatiotemporal block feature vector. Specifically, it includes: inputting the spatiotemporal block feature vectors into the incremental discriminant learning classification model; calculating the probability value of the spatiotemporal block feature vectors under preset categories through the hidden mapping layer and output layer of the incremental discriminant learning classification model; comparing the probability values ​​corresponding to the background class, the artificial disturbance class, and the mechanical disturbance class respectively to determine the disturbance category corresponding to the spatiotemporal block feature vectors, and using the disturbance category as the block-level recognition result of the spatiotemporal block feature vectors. The correction unit is used to make decision corrections for each block-level recognition result to obtain a corrected block-level recognition result. Specifically, it includes: when the perturbation category of the spatiotemporal block feature vector is the artificial perturbation class, if the probability value of the spatiotemporal block feature vector in the artificial perturbation class is less than a preset minimum artificial confidence level, then the perturbation category of the spatiotemporal block feature vector is corrected to the background class to obtain the corrected block-level recognition result; when the perturbation category of the spatiotemporal block feature vector is the mechanical perturbation class, if the probability value of the spatiotemporal block feature vector in the mechanical perturbation class is less than a preset minimum mechanical confidence level threshold .... The category is corrected to the background class to obtain the corrected block-level recognition result; when the probability value of the spatiotemporal block feature vector in the mechanical disturbance class is greater than or equal to the preset mechanical forced reversal threshold, and the difference between the probability value of the spatiotemporal block feature vector in the mechanical disturbance class and the probability value of the spatiotemporal block feature vector in the background class is greater than or equal to the preset mechanical relative advantage margin, and the probability value of the spatiotemporal block feature vector in the mechanical disturbance class is greater than or equal to the probability value of the spatiotemporal block feature vector in the artificial disturbance class, then the disturbance category of the spatiotemporal block feature vector is corrected to the mechanical disturbance class to obtain the corrected block-level recognition result; The segment extraction unit is used to determine the valid event segments of the distributed optical fiber based on all the corrected block-level identification results. The warning generation unit is used to perform vehicle diagonal texture filtering based on the effective event segment to obtain the warning result of the effective event segment.

8. An electronic device, characterized in that, include: Processor and memory, the memory being used to store computer programs; When the computer program is loaded by the processor, it causes the processor to execute the distributed optical fiber vibration event identification and early warning method as described in any one of claims 1-6.

Images