Optical fiber signal feature generation method and system combining dynamic noise reduction and event screening
By using adaptive filtering and two-dimensional time-frequency feature conversion, the signal processing problem of fiber optic vibration sensing technology in complex noise environments was solved, and high-precision, real-time threat event identification was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUXI HUIXIAN SENSING TECHNOLOGY CO LTD
- Filing Date
- 2026-03-13
- Publication Date
- 2026-06-19
AI Technical Summary
Existing fiber optic vibration sensing technology struggles to distinguish between real threats and interference in complex noise environments, resulting in a high false alarm rate. It also struggles to balance data processing efficiency with identification accuracy, and its effective signal feature representation is insufficient, thus affecting the practicality and accuracy of early warning systems.
An adaptive filtering process using a spatial sliding window strategy, combined with energy screening and spatial continuity judgment, is adopted to convert the data into a two-dimensional time-frequency feature map, generating high-quality feature data.
It effectively reduces noise interference, improves the signal-to-noise ratio, quickly locates suspected events, reduces computing resource consumption, improves recognition accuracy and system response real-time performance, and adapts to complex noise environments.
Smart Images
Figure CN122241351A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of fiber optic sensing and digital signal processing technology, specifically to a method and system for generating fiber optic signal features that combines dynamic noise reduction and event screening. Background Technology
[0002] With the continuous expansion of urban infrastructure, the safe operation of underground pipelines (including municipal water supply and drainage, gas pipelines, and long-distance oil and gas pipelines) faces severe challenges. Damage caused by third-party construction (such as excavator operations, manual digging, and heavy vehicle crushing) has become a major cause of pipeline damage, posing a serious threat to public safety and business operations.
[0003] Traditional pipeline safety protection methods, such as regular manual inspections and video surveillance, have significant limitations when dealing with long-distance, underground pipelines. Distributed Acoustic Sensing (DAS) technology, based on a phase-sensitive optical time-domain reflectometer (Φ-OTDR), has become an important technological direction in the field of pipeline safety early warning because it can utilize existing communication optical cables to achieve continuous, real-time, and distributed vibration monitoring along pipelines tens or even hundreds of kilometers long.
[0004] However, when moving DAS technology from the laboratory to large-scale engineering applications, a series of technical bottlenecks still need to be addressed, mainly in the following aspects:
[0005] (1) Weak resistance to complex noise interference and high false alarm rate. The actual pipeline environment is complex and variable, with a large number of non-threatening vibration interferences, such as highway traffic noise, instantaneous impact of vehicles running over manhole covers or stones, and natural wind and rain disturbances. The time-domain waveforms, frequency-domain characteristics, and even energy intensity of these interference signals are often highly similar to real threatening events such as excavator operations. Existing methods mostly use fixed filtering thresholds or simple energy judgment based on single-channel signals. Such methods not only have difficulty in stably distinguishing between real threats and interferences in complex and variable noise backgrounds, but also lack the ability to identify the spatial continuity of suspected events. For example, a certain isolated instantaneous impact signal (such as a heavy object falling) may be falsely alarmed due to energy exceeding the limit, while real excavation events usually produce continuous responses on multiple spatially adjacent sensing units. This lack of spatial correlation information further aggravates the false alarm rate of the system and seriously reduces the practicality and user trust of the early warning system.
[0006] (2) It is difficult to balance data processing efficiency and recognition accuracy. The amount of data generated by the DAS system can reach GB per second, which places extremely high demands on the real-time performance of the processing system. In pursuit of high recognition accuracy, some methods perform complex time-frequency analysis and feature calculations on all raw data, consuming a large amount of computing resources, resulting in early warning delays and losing the significance of real-time early warning. Conversely, simplified algorithms adopted to meet real-time requirements often sacrifice recognition accuracy and are difficult to effectively detect early or weak threat signals.
[0007] (3) Insufficient feature representation and recognition capabilities of effective signals. After noise reduction and preliminary screening, how to characterize potential threat signals directly affects the accuracy of the final pattern recognition. Most existing methods directly use one-dimensional time-domain signals or their simple statistical features (such as mean and variance) as the basis for recognition. Such features have low information density and limited discriminative power, making it difficult to support high-precision classification models (such as convolutional neural networks). Summary of the Invention
[0008] The technical problem to be solved by this application is that the existing technology is insufficient in processing the original optical fiber signal. This application provides a method and system for generating optical fiber signal features that combines dynamic noise reduction and event screening, which can solve the technical problem that effective event features are difficult to extract and are easily confused with interference signals in complex noise environments.
[0009] To achieve the above objectives, the first aspect of this application provides a method for generating fiber optic signal features by combining dynamic noise reduction and event filtering, comprising: S1, acquire the multi-channel raw time-domain vibration signal collected by the distributed optical fiber vibration sensing system, and analyze and reconstruct the raw time-domain vibration signal to obtain a signal matrix. One dimension of the signal matrix corresponds to the spatial channel, and the other dimension corresponds to the time sampling point. S2, The signal matrix is subjected to adaptive filtering based on local noise estimation using a spatial sliding window strategy to obtain a denoised signal matrix; S3, based on the noise reduction signal matrix, perform energy screening and spatial continuity judgment in the spatial channel dimension to locate suspected event channel segments that meet the energy threshold and spatial continuity length conditions; S4, convert the one-dimensional time-domain signal corresponding to the suspected event channel segment into a two-dimensional time-frequency feature map, perform standardization processing, and output feature data for pattern recognition.
[0010] Specifically, the spatial sliding window strategy in S2 is as follows: the signal matrix is divided into multiple windows in the spatial channel dimension, each window covers a preset number of adjacent spatial channels, and the spatial sliding window is moved in the spatial channel dimension according to a preset step size, so that the spatial sliding window covers each window in turn, and the signal matrix is traversed window by window in the spatial channel dimension.
[0011] Specifically, the adaptive filtering process based on local noise estimation for the signal matrix in S2 includes: Within each spatial sliding window, based on the energy of the one-dimensional time-domain signal corresponding to each spatial channel within the window, the spatial channel with the lowest energy ratio is selected as a local noise reference, and a noise template is constructed based on the spatial channel used as the local noise reference to characterize the noise characteristics of the spatial sliding window. Based on the noise template, adaptive filtering is performed on the one-dimensional time-domain signal corresponding to each spatial channel within the spatial sliding window.
[0012] Specifically, adaptive filtering employs the Wiener filtering algorithm, and its filter's frequency domain response... Power spectrum estimation of the one-dimensional time-domain signal corresponding to each spatial channel With the noise template A joint decision, expressed as: ; in, These are positive constants used to ensure numerical stability.
[0013] Specifically, S3 performs energy screening and spatial continuity judgment in the spatial channel dimension to locate suspected event channel segments that meet the energy threshold and spatial continuity length conditions. These segments include: The energy value of each spatial channel is calculated within the window range corresponding to the spatial sliding window, and an energy threshold is generated by combining the statistical characteristics of the energy value within the window. High-energy channels with energy not lower than the energy threshold are then selected. The high-energy channels are spatially continuous. Channel segments with a number of continuous channels not less than a preset minimum continuous length are retained to obtain suspected event channel segments.
[0014] Specifically, energy threshold The formula for calculation is:
[0015] in, This indicates the median operation. The preset sensitivity factor, For the first The energy value of each spatial channel This represents the number of spatial channels within the sliding window.
[0016] Specifically, in S4, converting the one-dimensional time-domain signal corresponding to the suspected event channel segment into a two-dimensional time-frequency feature map includes: For the one-dimensional time-domain signal corresponding to each channel in each suspected event channel segment obtained by screening, a short-time Fourier transform is performed and the power spectral density is calculated. The power spectral density is then mapped to the Mel frequency scale through a Mel filter bank and the logarithm is taken to generate a two-dimensional Mel spectrum.
[0017] A second aspect of this application provides a fiber optic signal feature generation system that combines dynamic noise reduction and event filtering, comprising: The signal acquisition module acquires multi-channel raw time-domain vibration signals collected by the distributed optical fiber vibration sensing system, and performs analytical reconstruction on the raw time-domain vibration signals to obtain a signal matrix. One dimension of the signal matrix corresponds to the spatial channel, and the other dimension corresponds to the time sampling point. The dynamic adaptive noise reduction module uses a spatial sliding window strategy to perform adaptive filtering on the signal matrix based on local noise estimation to obtain a noise-reduced signal matrix. The event filtering module, based on the noise-reduced signal matrix, performs energy filtering and spatial continuity judgment in the spatial channel dimension to locate suspected event channel segments that meet the energy threshold and spatial continuity length conditions. The feature extraction module converts the one-dimensional time-domain signal corresponding to the suspected event channel segment into a two-dimensional time-frequency feature map, performs standardization processing, and outputs feature data for pattern recognition.
[0018] The above technical solution achieves the following technical effects: 1. It can estimate and suppress local noise characteristics in different sections along the pipeline in real time, rather than using a globally fixed threshold. This effectively improves the signal-to-noise ratio of the original vibration signal, reduces the interference of noise on subsequent event identification from the source, and lays a solid foundation for high-precision analysis.
[0019] 2. Through a two-stage rapid screening mechanism of initial energy screening and spatial continuity verification, a very small number of suspected event channels can be quickly located from massive amounts of data, filtering out more than 99% of background data. This allows subsequent computationally intensive feature transformation and model recognition to process only a very small amount of data, greatly saving computing resources and ensuring the real-time response of the entire monitoring system.
[0020] 3. The one-dimensional time-domain signal is converted into a two-dimensional Mel-spectrum image containing rich time-frequency information. This feature form has high information density and strong discriminative power, which is more in line with the cognitive characteristics of human-machine collaboration and can highlight subtle differences in event features. The high-quality feature vectors generated after standardization provide high-quality input for various deep learning classification models, thereby significantly improving the overall accuracy and reliability of the system in identifying different types of threat events (such as excavator operations and manual digging). Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 This is an overall flowchart of the optical fiber signal feature generation method provided in the embodiments of the present invention.
[0023] Figure 2 This is a schematic diagram illustrating the principle of the dynamic adaptive noise reduction module in this embodiment of the invention, which uses a spatial sliding window for local noise estimation and filtering.
[0024] Figure 3 This is a schematic diagram illustrating the principle of the event filtering module performing initial energy screening and spatial continuity verification within a spatial sliding window in this embodiment of the invention.
[0025] Figure 4 This is a schematic diagram of the feature conversion process in an embodiment of the present invention, which converts a one-dimensional time-domain vibration signal into a two-dimensional Mel spectrum. Detailed Implementation
[0026] To facilitate understanding of this application, the following description will be more comprehensive and detailed in conjunction with the accompanying drawings and preferred embodiments, but the scope of protection of this application is not limited to the following specific embodiments.
[0027] Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the scope of this application.
[0028] In one embodiment of this application, a method for generating fiber optic signal features by combining dynamic noise reduction and event filtering is provided, comprising the following steps: S1, acquire the multi-channel original time-domain vibration signal of the distributed optical fiber vibration sensing system, and analyze and reconstruct the original time-domain vibration signal to obtain the signal matrix; In this step, the multi-channel original time-domain vibration signal is a vibration sampling sequence collected by the distributed optical fiber vibration sensing system along different spatial channels of the optical fiber. It is analyzed to obtain sampling parameters such as sampling rate, number of sampling points, and number of channels. The time-domain sampling sequence of each spatial channel is reconstructed according to the channel order so that one dimension of the signal matrix corresponds to the spatial channel and the other dimension corresponds to the time sampling point. S2, The signal matrix is subjected to adaptive filtering based on local noise estimation using a spatial sliding window strategy to obtain a denoised signal matrix; In this step, the spatial sliding window moves along the spatial channel dimension and covers the signals of adjacent channels. The noise level is locally estimated within the window, and the filtering strength or filtering parameters are adaptively adjusted based on the local noise estimation results. The signal matrix is filtered to suppress noise and output a noise-reduced signal matrix. S3, based on the noise reduction signal matrix, performs energy screening and spatial continuity judgment in the spatial channel dimension to locate suspected event channel segments that meet the energy threshold and spatial continuity length conditions; S4, convert the one-dimensional time-domain signal corresponding to the suspected event channel segment into a two-dimensional time-frequency feature map, perform standardization processing, and output feature data for pattern recognition.
[0029] The method in this embodiment achieves adaptive noise reduction through local noise estimation under a spatial sliding window, which can maintain stable noise suppression capability when the noise level changes at different spatial locations. Based on the noise reduction, energy screening and spatial continuity judgment are introduced, which can effectively eliminate isolated high-energy noise points and focus on spatially continuous suspected event channel segments, thereby reducing false detections and invalid calculations. Finally, the one-dimensional time-domain signal of the suspected event is converted into a standardized two-dimensional time-frequency feature map, which helps to highlight the time-frequency texture differences of the event and improve the accuracy of subsequent pattern recognition.
[0030] In one embodiment, the spatial sliding window strategy in S2 specifically involves windowing the signal matrix along the spatial channel dimension and traversing the signal matrix window by window using a spatial sliding window. That is, the signal matrix is divided into multiple windows along the spatial channel dimension, each window covering a preset number of adjacent spatial channels. The spatial sliding window moves along the spatial channel dimension according to a preset step size, so that the spatial sliding window sequentially covers each window, thereby achieving window-by-window traversal of the signal matrix along the spatial channel dimension.
[0031] In this embodiment, the adaptive filtering process based on local noise estimation of the signal matrix in S2 specifically includes: Within each spatial sliding window, a subset of channels are selected as local noise references based on the energy of each channel signal within the window, and a noise template is constructed based on the local noise references to characterize the noise characteristics of that spatial sliding window. Adaptive filtering is performed on the channel signals within the sliding window of the space based on the noise template.
[0032] In one embodiment, Adaptive filtering employs the Wiener filtering algorithm, and its filter's frequency domain response... Power spectrum estimation of the one-dimensional time-domain signal corresponding to each spatial channel With the noise template A joint decision, expressed as: ; in, These are positive constants used to ensure numerical stability.
[0033] In one embodiment, S3 involves energy screening and spatial continuity assessment along the spatial channel dimension to locate suspected event channel segments that meet the energy threshold and spatial continuity length conditions. Specifically, this includes: Within the window range corresponding to the spatial sliding window, the energy value of each channel is calculated, and an energy threshold is generated based on the statistical characteristics of the energy values within the window. High-energy channels with energy values not lower than the energy threshold are then selected. Specifically, within the spatial channel range covered by the spatial sliding window, the energy value is calculated for each spatial channel within the window. The energy value is used to characterize the vibration intensity of the channel within the window range. An energy threshold is generated based on the statistical characteristics of the energy values of each channel within the window. The energy threshold is adjusted according to the energy distribution within the window, thereby identifying spatial channels with energy values not lower than the energy threshold as high-energy channels. The spatial continuity of high-energy channels is assessed, and channel segments with a number of consecutive channels not less than a preset minimum continuous length are retained to obtain suspected event channel segments. Specifically, high-energy channels are checked for adjacency according to their natural spatial order. High-energy channels that are adjacent or consecutive in the spatial channel dimension are merged into channel segments, and the number of consecutive channels contained in each channel segment is calculated. Channel segments with a number of consecutive channels not less than a preset minimum continuous length are retained and output as suspected event channel segments.
[0034] This step allows for the rapid identification of a small number of truly concerning potential threat events, while filtering out the vast majority of background noise and isolated interference. To accommodate the spatial non-uniformity of background noise levels along long-distance pipelines, event filtering is performed independently within the same spatial sliding window as dynamic noise reduction.
[0035] In one embodiment, S4, converting the one-dimensional time-domain signal corresponding to the suspected event channel segment into a two-dimensional time-frequency feature map, specifically includes: For the one-dimensional time-domain signal corresponding to each channel in each suspected event channel segment obtained by screening, short-time Fourier transform is performed and power spectral density is calculated. The power spectral density is then mapped to the Mel frequency scale through the Mel filter bank and the logarithm is taken to generate a two-dimensional Mel spectrum. Specifically, this step involves extracting the corresponding one-dimensional time-domain signal from each spatial channel within each suspected event channel segment. The one-dimensional time-domain signal is then processed into frames according to a preset window function, frame length, and frame shift, and a short-time Fourier transform is performed to obtain a time-frequency representation. Based on this, the power spectral density of each time-frequency unit is calculated. Then, a Mel filter bank is used to weight and accumulate the power spectral density in the frequency dimension, mapping the frequency axis from a linear frequency scale to a Mel frequency scale. The logarithm of the mapped result is then taken to generate a two-dimensional Mel spectrum for characterizing the time-Mel frequency-energy distribution, which is output as a two-dimensional time-frequency feature map.
[0036] This step can enhance the separability of different event types in terms of spectral morphology and suppress the impact of amplitude scale differences, thereby improving the accuracy of fiber optic vibration event identification.
[0037] In one embodiment, this application also provides an optical fiber signal feature generation system combining dynamic noise reduction and event filtering, comprising: The signal acquisition module acquires multi-channel raw time-domain vibration signals collected by the distributed optical fiber vibration sensing system, and performs analytical reconstruction on the raw time-domain vibration signals to obtain a signal matrix. One dimension of the signal matrix corresponds to the spatial channel, and the other dimension corresponds to the time sampling point. The dynamic adaptive noise reduction module uses a spatial sliding window strategy to perform adaptive filtering on the signal matrix based on local noise estimation to obtain a noise-reduced signal matrix. The event filtering module, based on the noise-reduced signal matrix, performs energy filtering and spatial continuity judgment in the spatial channel dimension to locate suspected event channel segments that meet the energy threshold and spatial continuity length conditions. The feature extraction module converts the one-dimensional time-domain signal corresponding to the suspected event channel segment into a two-dimensional time-frequency feature map, performs standardization processing, and outputs feature data for pattern recognition.
[0038] The system includes a processor and a memory. The aforementioned modules are all stored in the memory as program units, and the processor executes the aforementioned program modules stored in the memory to implement the corresponding functions.
[0039] like Figure 1 , 2 As shown, in one embodiment, each module can be initialized first; The dynamic adaptive noise reduction module has a pre-defined uniform window length. (e.g., 80 channels) and overlap length (e.g., 10 channels) and sliding step size (e.g., 70 channels); The event filtering module sets the energy threshold sensitivity factor (e.g., 4.0) and the minimum number of continuous channels for spatial continuity determination (e.g., 3). The feature extraction module sets parameters such as the sampling rate (e.g., 1000Hz), frame length of the short-time Fourier transform (e.g., 512), frame shift (e.g., 128), and number of Mel filters (e.g., 64) required for time-frequency conversion.
[0040] When a new raw vibration data file (such as a .dat file) is received, the subsequent control process is triggered; The system reads the raw vibration data file, parses the file header information, obtains the sampling frequency, number of spatial channels, and number of time sampling points, and parses the data file by channel to reconstruct a two-dimensional signal matrix. ,in Represents the total number of channels. This represents the number of time sampling points for each channel.
[0041] With length as A spatial sliding window (e.g., W=80) with a sliding step size of... (its value is) Slide along the channel dimension. For the first... Each window covers a channel index range of [number]. .
[0042] For the current window The one-dimensional time-domain signal (or simply channel signal) of each channel (spatial channel) is used to calculate the signal energy of each channel. .
[0043] Select the one with the lowest energy (For example The channel is used as a local noise reference.
[0044] The power spectrum of the one-dimensional time-domain signal of the channel used as a local noise reference is calculated by performing a Short-Time Fourier Transform (STFT). Then, the power spectrum of all channels used as local noise references is arithmetically averaged to obtain a noise template representing the local noise characteristics of the window. .
[0045] For each channel signal within the current window Using the obtained noise template Noise reduction is performed. One specific implementation uses frequency-domain Wiener filtering, whose frequency domain response... Power spectrum estimation from this channel signal Determined in conjunction with the noise template: ; in, It is a small positive number used to avoid division by zero errors.
[0046] Specifically, the channel signal STFT spectrum and Multiplying the two signals and then performing the inverse STFT yields the noise-reduced signal for that channel. This represents the one-dimensional time-domain signal after denoising. By iterating through all channels of this window, the denoised signal matrix is ultimately output. .
[0047] like Figure 3 As shown, for the noise reduction signal matrix The signal in, i.e., the first A sliding window covering the channel index ) within The noise-reduced signal Calculate the total energy of each spatial channel in the denoised signal matrix. For the ... Each spatial channel, its energy .
[0048] Based on the energy value set of all channels Dynamically calculate an energy threshold for screening. This method does not rely on a fixed threshold and can adapt to background noise levels at different times or locations. A preferred dynamic threshold calculation method is as follows:
[0049] in, This indicates the median operation. For the preset sensitivity factor (e.g.) Finally, satisfy all The channel is labeled as a "high-energy channel".
[0050] After obtaining the set of high-energy channels, it is necessary to further determine whether they constitute suspected event segments with spatial continuity in order to filter out isolated transient disturbances (such as heavy objects falling). The specific method is as follows: In the high-energy channel set, identify all channel segments that are contiguous on the spatial index.
[0051] For each consecutive channel segment, check the number of channels it contains. .
[0052] Only when At that time, the channel segment was identified as a valid "suspected incident channel segment". Among them, The preset minimum number of consecutive channels (e.g.) ).
[0053] This step ensures that only vibration signals that have a certain spatial continuity and are more likely to be generated by actual excavation or destructive activities are retained for subsequent depth analysis.
[0054] Repeat the above steps until all channels have been filtered. Then, integrate all filtered "suspected event channel segments" (including the index of each channel and its corresponding denoised signal data) as output. For suspected event channel segments obtained from adjacent overlapping windows that overlap in channel index, merge them before outputting. This output will serve as input to the subsequent feature extraction module, thereby reducing the data processing workload from all channels (…). By compressing (number of) channels into a few consecutive channel segments, the system efficiency is greatly improved.
[0055] By performing feature enhancement and formatting on the one-dimensional time-domain signal corresponding to each "suspected event channel segment" output by the event filtering module, it is converted into a two-dimensional feature map with higher information density and stronger discriminative power for use by subsequent pattern recognition models.
[0056] like Figure 4 As shown, for each denoised one-dimensional time-domain signal in the suspected event channel segment... First, perform a short-time Fourier transform (STFT) on it to obtain its joint time-frequency representation:
[0057] in, For window functions (such as Hamming windows). The length of the Fast Fourier Transform (FFT) is... This refers to frame shift (hop length). For intra-frame sampling point index, The base of the natural constant, The imaginary unit, For time frame indexing, This is the frequency index. Subsequently, its power spectral density estimate is calculated:
[0058] To simulate the characteristics of human hearing and identify more discriminative low-frequency information, the power spectrum is... Nonlinear mapping is performed using a set of Mel filters. Mel frequency. With linear frequency The conversion relationship is as follows:
[0059] Mel filter bank includes A triangular filter (e.g.) Power spectrum With the Mel filter Perform frequency domain convolution to obtain the energy at the Mel scale:
[0060] To compress the dynamic range and improve the robustness of the feature to amplitude variations, the logarithm of the Mel energy is taken:
[0061] in, It is a small positive constant used to ensure numerical stability. The result is... This is a two-dimensional Mel spectrogram, whose dimension is the number of Mel filters. Time frame count.
[0062] To eliminate differences in units and distributions among different samples and ensure that the feature data meets the input requirements of the model, the generated Mel spectrogram is standardized. The mean used for standardization is... and standard deviation The parameters are all pre-calculated in the training dataset:
[0063] in, Represents the mathematical expectation. During online processing, for each Mel band... Perform the following calculations independently:
[0064] The standardized two-dimensional feature matrix It is packaged with its corresponding metadata (such as channel number and timestamp) to form the final feature data that can be directly input into classification models such as convolutional neural networks.
[0065] The aforementioned models (such as convolutional neural networks) are pre-trained classification models. Their training data is constructed in the following way: vibration signal segments of various events (such as "background noise", "manual digging", "excavator operation") are collected and labeled manually, and then directly processed by short-time Fourier transform, Mel filter bank mapping and standardization to generate standardized Mel spectrograms as input features, and the corresponding categories are the output labels.
[0066] The specific implementation method for detecting excavator operation signals is as follows: Once a vibration signal is detected, the raw data file (e.g., a .dat file) containing the vibration signal from the excavator operation is loaded into the external damage event recognition algorithm system.
[0067] The system parses the .dat file and extracts its information, including the collection timestamp (e.g., 2026-01-08 18:00:00), sampling frequency (e.g., 1000Hz), number of spatial sampling points (e.g., corresponding to 10 kilometers, a total of 10,000 channels), and number of time sampling points (e.g., 1,000 samples per second).
[0068] Based on the parsed file information, the .dat file is divided and processed according to channels, and the data is reconstructed into a two-dimensional signal matrix. The rows represent channels, and the columns represent time sampling points.
[0069] Dynamic noise reduction is performed. The system employs a spatial sliding window strategy for noise reduction, initializing a window length of... The window, with sliding steps (i.e., an overlap length of 10 channels) slides along the channel dimension. For each window (e.g., a window covering channels 5000 to 5079), calculate the energy of all channels within the window.
[0070] Select the 20% of channels with the lowest energy (e.g., 16 channels) as the local noise reference for the current window, perform short-time Fourier transform and amplitude square calculation on these channels, and calculate their average to obtain a noise template specific to this window.
[0071] Using a noise template to analyze the signal of each channel within the window Perform Wiener filtering and calculate the denoised signal. Traverse all windows to obtain the global denoised signal matrix. This process effectively suppressed steady-state background interference such as vehicle noise in various regions.
[0072] To accommodate varying noise levels in long-distance pipelines, event filtering is performed independently within the same spatial sliding window as noise reduction. The energy of all 80 noise-reduced channels is calculated, the median of these energies is taken, and multiplied by a preset sensitivity factor (e.g., 4.0). This dynamically generates an energy threshold T that applies only to this window. Channels with energy levels not lower than T within the window are marked as high-energy channels. For example, if 24 high-energy channels exceeding the threshold are selected, the remaining channels are ignored.
[0073] These high-energy channels are screened, and only channel segments whose length exceeds a preset minimum number of consecutive channels (e.g., 3) are retained. These segments are identified as "locally suspected event segments." After all channels have undergone dynamic noise reduction and event filtering, the system integrates all identified "locally suspected event segments": if two "locally suspected event segments" overlap in their channel indices, they are merged into a longer consecutive segment. Finally, the system outputs all consecutive segments (including their channel indices and the corresponding denoised signal data) as input to the feature extraction module. For example, it outputs signal data from 5005-5010 and 5012-5015.
[0074] For the signals of the globally suspected event channel segments finally selected above, time-frequency feature transformation is performed on each one. For the one-dimensional signal of each channel... STFT is performed and the power spectrum is calculated. Then, a two-dimensional Mel spectrum of size 64 (Mel band) × 36 (time frame) is generated through a set of 64 Mel filter banks. The spectrum is then standardized using pre-calculated mean and standard deviation parameters.
[0075] The generated standardized Mel-ray spectrogram, along with its metadata (such as channel number 5005 and timestamp), is packaged into a structured feature data object and input into a pre-trained pattern recognition model (such as a convolutional neural network). The model uses these features to determine that the event is an "excavator operation" and locates the threat near channel 5010. If the model determines it to be background noise, no alarm is triggered.
[0076] The above are merely preferred embodiments of this application. It should be noted that this application is not limited to the above embodiments. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should also be considered within the scope of protection of this application.
Claims
1. A method of optical fiber signal feature generation combining dynamic noise reduction and event screening, the method comprising: include: S1, acquire the multi-channel raw time-domain vibration signal collected by the distributed optical fiber vibration sensing system, and analyze and reconstruct the raw time-domain vibration signal to obtain a signal matrix. One dimension of the signal matrix corresponds to the spatial channel, and the other dimension corresponds to the time sampling point. S2, The signal matrix is subjected to adaptive filtering based on local noise estimation using a spatial sliding window strategy to obtain a denoised signal matrix; S3, based on the noise reduction signal matrix, perform energy screening and spatial continuity judgment in the spatial channel dimension to locate suspected event channel segments that meet the energy threshold and spatial continuity length conditions; S4, convert the one-dimensional time-domain signal corresponding to the suspected event channel segment into a two-dimensional time-frequency feature map, perform standardization processing, and output feature data for pattern recognition.
2. The method of claim 1, wherein, The spatial sliding window strategy in S2 is as follows: the signal matrix is divided into multiple windows in the spatial channel dimension, each window covers a preset number of adjacent spatial channels, and the spatial sliding window is moved in the spatial channel dimension according to a preset step size so that the spatial sliding window covers each window in turn, thereby realizing window-by-window traversal processing of the signal matrix in the spatial channel dimension.
3. The method of claim 2, wherein, The adaptive filtering process based on local noise estimation of the signal matrix in S2 specifically includes: Within each spatial sliding window, based on the energy of the one-dimensional time-domain signal corresponding to each spatial channel within the window, the spatial channel with the lowest energy ratio is selected as a local noise reference, and a noise template is constructed based on the spatial channel used as the local noise reference to characterize the noise characteristics of the spatial sliding window. Based on the noise template, adaptive filtering is performed on the one-dimensional time-domain signal corresponding to each spatial channel within the spatial sliding window.
4. The fiber optic signal feature generation method combining dynamic noise reduction and event screening according to claim 3, characterized in that, The adaptive filtering employs the Wiener filtering algorithm, and its filter's frequency domain response... Power spectrum estimation of the one-dimensional time-domain signal corresponding to each spatial channel With the noise template A joint decision, expressed as: ； in, These are positive constants used to ensure numerical stability.
5. The fiber optic signal feature generation method combining dynamic noise reduction and event screening according to claim 1, characterized in that, In step S3, energy screening and spatial continuity judgment are performed along the spatial channel dimension to locate suspected event channel segments that meet the energy threshold and spatial continuity length conditions. Specifically, this includes: The energy value of each spatial channel is calculated within the window range corresponding to the spatial sliding window, and an energy threshold is generated by combining the statistical characteristics of the energy value within the window. High-energy channels with energy not lower than the energy threshold are then selected. The high-energy channels are spatially continuous. Channel segments with a number of continuous channels not less than a preset minimum continuous length are retained to obtain suspected event channel segments.
6. The fiber optic signal feature generation method combining dynamic noise reduction and event screening according to claim 5, characterized in that, The energy threshold The formula for calculation is: in, This indicates the median operation. The preset sensitivity factor, For the first The energy value of each spatial channel This represents the number of spatial channels within the sliding window.
7. The fiber optic signal feature generation method combining dynamic noise reduction and event screening according to claim 1, characterized in that, The step S4, which converts the one-dimensional time-domain signal corresponding to the suspected event channel segment into a two-dimensional time-frequency feature map, specifically includes: For the one-dimensional time-domain signal corresponding to each channel in each suspected event channel segment obtained by screening, a short-time Fourier transform is performed and the power spectral density is calculated. The power spectral density is then mapped to the Mel frequency scale through a Mel filter bank and the logarithm is taken to generate a two-dimensional Mel spectrum.
8. A fiber optic signal feature generation system combining dynamic noise reduction and event filtering, characterized in that, include: The signal acquisition module acquires multi-channel raw time-domain vibration signals collected by the distributed optical fiber vibration sensing system, and performs analytical reconstruction on the raw time-domain vibration signals to obtain a signal matrix. One dimension of the signal matrix corresponds to the spatial channel, and the other dimension corresponds to the time sampling point. The dynamic adaptive noise reduction module uses a spatial sliding window strategy to perform adaptive filtering on the signal matrix based on local noise estimation to obtain a noise-reduced signal matrix. The event filtering module, based on the noise-reduced signal matrix, performs energy filtering and spatial continuity judgment in the spatial channel dimension to locate suspected event channel segments that meet the energy threshold and spatial continuity length conditions. The feature extraction module converts the one-dimensional time-domain signal corresponding to the suspected event channel segment into a two-dimensional time-frequency feature map, performs standardization processing, and outputs feature data for pattern recognition.