Distributed acoustic sensing signal reconstruction method and device, storage medium, program product

Abstract

The application relates to a distributed acoustic sensing signal reconstruction method and device, a storage medium and a program product, which comprises the following steps: collecting low-resolution original signals of a distributed acoustic sensing system; sequentially performing normalization processing, detrending processing, space-time slicing and up-sampling processing on the low-resolution original signals, and converting the signals into tensor data; inputting the tensor data into a pre-trained ResUNet++ deconvolution neural network model for forward inference, and outputting a super-resolution distributed acoustic sensing signal; sequentially performing amplitude and phase calibration, vibration event space positioning optimization and visualization processing on the super-resolution distributed acoustic sensing signal, and outputting position, quantity, waveform and range information of fiber vibration events. The application can break through the physical limit of spatial resolution without changing hardware, and has the advantages of high resolution, strong robustness and controllable cost.

Description

Technical Field

[0001] This application belongs to the field of distributed optical fiber sensing technology, specifically relating to a distributed acoustic sensing signal reconstruction method and device, storage medium, and program product. Background Technology

[0002] Distributed Acoustic Sensing (DAS) monitors the phase changes of backscattered Rayleigh light in optical fibers and is widely used in structural health monitoring, traffic monitoring, and geological activity monitoring. Spatial resolution is a key parameter of DAS systems, typically determined by the laser pulse width, and represents the minimum distance at which the system can distinguish two adjacent events.

[0003] While pulse compression and interferometer-assisted methods have shown some effectiveness in improving spatial resolution, they require hardware modifications, increasing system cost and complexity. In recent years, deconvolution methods based on neural networks have been introduced into DAS signal processing. Dong et al. proposed network architectures such as FSRCNN, and Huang et al. proposed ConverseNet, which initially achieved resolution improvement. However, existing methods still have the following shortcomings: (1) The network architecture is simple and lacks multi-scale feature extraction capabilities; (2) The algorithm has poor robustness and its performance degrades significantly in complex scenarios; (3) The pure data-driven approach lacks physical constraints, and the output signal is prone to distortion. In addition, there is a traditional trade-off between spatial resolution and sensing distance in DAS systems, which affects economic efficiency.

[0004] Therefore, the present invention provides a method for reconstructing distributed acoustic sensing signals with super spatial resolution. Summary of the Invention

[0005] The purpose of this application is to propose a distributed acoustic sensor signal reconstruction method based on DAS, which can break through the physical resolution limit determined by pulse width and realize distributed acoustic sensor signal reconstruction with super spatial resolution without changing the hardware conditions.

[0006] To achieve the above objectives, embodiments of this application provide a distributed acoustic sensor signal reconstruction method, comprising: Acquire low-resolution raw signals from a distributed acoustic sensing system; The low-resolution original signal is sequentially normalized, detrended, spatiotemporally sliced, and upsampled to be converted into tensor data. The tensor data is input into a pre-trained ResUNet++ deconvolutional neural network model for forward inference, and outputs a super-resolution distributed acoustic sensing signal; the pre-trained model is used to map the low-resolution distributed acoustic sensing signal into a super-resolution distributed acoustic sensing signal. The super-resolution distributed acoustic sensing signal is sequentially subjected to amplitude and phase calibration, vibration event spatial positioning optimization, and visualization processing to output the location, quantity, waveform, and range information of vibration events along the optical fiber.

[0007] Preferably, the upsampling process adjusts the signal spatial sampling interval to a preset target value; the detrending process is used to eliminate the DC component and trend term in the signal; and the spatiotemporal slicing is used to convert the two-dimensional spatiotemporal signal into a signal segment of fixed size.

[0008] Preferably, the visualization process includes generating a waterfall plot and a vibration localization curve. The waterfall plot is used to display the intensity or phase distribution of the super-resolution distributed acoustic sensing signal in a two-dimensional time-space plane, and the vibration localization curve is used to display the distribution curve of the vibration signal amplitude or energy along the optical fiber. The vibration event spatial localization optimization adopts a strategy combining peak detection and threshold screening. Specifically, an amplitude threshold is set, and the peak positions of the super-resolution distributed acoustic sensing signal with amplitudes exceeding the threshold are determined as the locations where vibration events occur, while noise signals below the threshold are removed.

[0009] Preferably, the ResUNet++ deconvolutional neural network model sequentially includes a Stem module, a ResNet module, a dilated spatial pyramid pooling module, a decoding module, and an output module. The Stem module is used to perform initial convolutional feature extraction on the input tensor. The ResNet module is used to extract deep signal features through residual connections and attention mechanisms. The dilated spatial pyramid pooling module is used to capture multi-scale contextual information through parallel convolutional layers with different dilation rates. The decoding module is used to restore the signal spatial resolution step by step through upsampling and attention mechanisms. The output module is used to aggregate multi-scale features and output a super-resolution distributed acoustic sensing signal.

[0010] Preferably, the low-resolution original signal is a backscattered Rayleigh light signal collected by the distributed acoustic sensing system after emitting a laser pulse into the sensing fiber with a predetermined pulse width, and obtained after interferometric detection and phase demodulation; the spatial resolution of the super-resolution distributed acoustic sensing signal is less than 2m, and it can distinguish multiple adjacent vibration events.

[0011] Preferably, the pre-trained ResUNet++ deconvolutional neural network model is obtained in the following way: Construct a ResUNet++ deconvolutional neural network model; Obtain a training dataset, which contains multiple sample pairs consisting of low-resolution distributed acoustic sensing signals and corresponding high-resolution label signals; The ResUNet++ deconvolutional neural network model is trained using the low-resolution distributed acoustic sensing signal as input and the corresponding high-resolution label signal as output. The loss function during training includes a data fidelity term and a physical constraint term. The physical constraint term is constructed based on the convolutional degradation model of the distributed acoustic sensing system, enabling the network to fit the inverse mapping of the fiber optic response function.

[0012] Preferably, in the training dataset, the low-resolution distributed acoustic sensing signal of each sample pair corresponds to the first spatial resolution, and the high-resolution tag signal corresponds to the second spatial resolution, wherein the second spatial resolution is higher than the first spatial resolution; the training dataset is generated through system simulation and covers multiple application scenarios with different numbers of vibration events, and the waveform, position, and length of each vibration event are randomized; the convolutional degradation model is used to simulate the signal degradation process caused by the superposition of noise after the convolution of the probe pulse time profile and the fiber response function.

[0013] This application also provides a distributed acoustic sensor signal reconstruction system, including: The signal acquisition module is used to acquire low-resolution raw signals from the distributed acoustic sensing system. The preprocessing module is used to perform normalization, detrending, spatiotemporal slicing and upsampling on the low-resolution raw signal in sequence, and convert it into tensor data. The super-resolution reconstruction module is loaded with a pre-trained ResUNet++ deconvolutional neural network model, which is used to perform forward inference on the tensor data and output a super-resolution distributed acoustic sensing signal; the pre-trained model is used to map the low-resolution distributed acoustic sensing signal into a super-resolution distributed acoustic sensing signal. The post-processing output module is used to sequentially perform amplitude and phase calibration, vibration event spatial positioning optimization, and visualization processing on the super-resolution distributed acoustic sensing signal, and output the location, quantity, waveform, and range information of vibration events along the optical fiber.

[0014] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the distributed acoustic sensor signal reconstruction method as described in this application.

[0015] This application also provides a computer program product, including a computer program or computer executable instructions, which, when executed by a processor, implement the distributed acoustic sensor signal reconstruction method as described in this application.

[0016] The embodiments of this application have the following beneficial effects: The super spatial resolution distributed acoustic sensor signal reconstruction method provided in this application uses a pre-trained ResUNet++ deconvolutional neural network model to perform forward inference on low-resolution DAS signals. Without changing the laser pulse width or increasing hardware complexity, it can improve the system spatial resolution from the traditional 10m to below 2m, breaking through the physical resolution limit determined by the pulse width. At the same time, this application embodiment can adapt to DAS signals with different pulse widths and sampling intervals, has strong robustness, can maintain high reconstruction accuracy in complex environments, and requires no hardware modification. Resolution improvement can be achieved only through back-end algorithms, balancing high performance and low cost, and is easy to implement in engineering. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings required in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a distributed acoustic sensor signal reconstruction method according to an embodiment of this application. Detailed Implementation

[0019] The detailed description of the accompanying drawings is intended to illustrate some embodiments of this application and is not intended to represent only the forms in which this application can be implemented. It should be understood that the same or equivalent functions can be accomplished by different embodiments intended to be included within the spirit and scope of this application.

[0020] like Figure 1 As shown, one embodiment of this application provides a distributed acoustic sensor signal reconstruction method, including the following steps: Step S10: Acquire the low-resolution raw signal from the distributed acoustic sensing system.

[0021] Specifically, when the DAS system is operating, a narrow-pulse laser source emits laser pulses into the sensing fiber. As the laser pulses propagate through the fiber, they interact with microscopic defects within the fiber, generating backscattered Rayleigh light. When external vibrations act on the fiber, they cause localized strain, which modulates the phase of the scattered light. The interferometric detection module converts the received scattered light signal into an electrical signal, and the phase demodulation module uses a phase difference algorithm to demodulate the phase signal corresponding to the vibration from the electrical signal. This signal is the low-resolution DAS raw signal. It should be noted that this low resolution refers to its spatial resolution being limited by the laser pulse width; for example, a 100ns pulse width corresponds to a spatial resolution of 10m.

[0022] Step S20: The low-resolution original signal is sequentially normalized, detrended, spatiotemporally sliced, and upsampled to convert it into tensor data.

[0023] Specifically, normalization unifies the signal amplitude to a uniform scale to avoid the impact of data of different magnitudes on network training and inference. Detrending eliminates DC components and trend terms in the signal, which typically originate from baseline drift within the system itself, not from vibrational events; failure to eliminate these components can interfere with subsequent processing. Spatiotemporal slicing cuts the two-dimensional spatiotemporal signal into fixed-size segments according to a preset window size to accommodate the fixed input size requirements of the neural network. Upsampling adjusts the spatial sampling interval of the signal to a preset target value, aiming to match the spatial resolution of the input signal with the data used during network training. In one embodiment, this preset target value is 0.125m.

[0024] Step S30: Input the tensor data into the pre-trained ResUNet++ deconvolutional neural network model for forward inference and output the super-resolution distributed acoustic sensing signal.

[0025] Specifically, the ResUNet++ deconvolutional neural network model used is a pre-trained deep learning model whose function is to map low-resolution DAS signals to super-resolution DAS signals. It is important to understand that "super-resolution" here refers to the output signal having a better spatial resolution than the input signal. For example, if the input signal has a resolution of 10m, the output signal can achieve a resolution of less than 2m.

[0026] More specifically, the ResUNet++ deconvolutional neural network model consists of a Stem module, a ResNet module, a Spatial Pyramid Pooling (ASPP) module, a decoding module, and an output module, and adopts a symmetrical encoder-decoder architecture.

[0027] The Stem module, located at the front end of the network, performs initial convolutional feature extraction on the input tensor, transforming the original signal into a feature space. This module typically contains several convolutional layers, responsible for extracting shallow features of the signal, such as basic information like signal edges and trends.

[0028] The ResNet module, connected after the Stem module, forms the core of the encoder. This module extracts deep signal features through residual connections and an attention mechanism. The residual connections allow gradients to propagate effectively within the deep network, preventing degradation as the network deepens. The attention mechanism allows the network to automatically focus on information-rich regions of the signal, such as signal changes near the vibration location, while suppressing responses from irrelevant regions. Through this module's processing, the network can extract deep abstract features from the DAS signal that are highly correlated with the vibration event.

[0029] The ASPP module connects the encoder and decoder, following the ResNet module and preceding the decoder, acting as a bridge between them. This module processes the feature map using multiple parallel convolutional layers with different dilation rates, then fuses the outputs of each layer. Convolutional layers with different dilation rates have different receptive fields; low-dilation layers focus on local details, while high-dilation layers focus on global context. By fusing multi-scale information, the ASPP module enables the network to simultaneously perceive both local vibration features and the overall distribution of vibrations, which is crucial for distinguishing between adjacent vibration events.

[0030] The decoding module is the core of the decoder, used to progressively restore the spatial resolution of the signal through upsampling and attention mechanisms. While the encoder compresses the feature map size step-by-step through downsampling, the decoding module progressively restores the feature map size through upsampling, thus restoring the output signal to the original input's temporal-spatial dimensions. In this process, an attention mechanism is also introduced to enhance the recovery of important features.

[0031] The output module is the last layer of the network. It is used to aggregate the multi-scale features output by the decoding module and map the feature map into the final output signal, namely the super-resolution distributed acoustic sensing signal, through convolution operation.

[0032] The theoretical basis for this network's ability to achieve super-resolution reconstruction lies in the fact that the signal formation process of the DAS system can be modeled as the convolution of the probe pulse and the fiber response function. The observed Rayleigh backscattered signal can be expressed as:

[0033] Where p(t) is the probe pulse time profile, h(t) is the fiber response function, and n(t) is the noise. The spatial resolution of a traditional DAS system is determined by the pulse width, and its physical resolution limit is:

[0034] Where τ is the pulse duration, c is the speed of light in vacuum, and n is the refractive index of the optical fiber. This formula shows that, under the condition of a fixed pulse width, the system theoretically has an insurmountable lower limit of resolution. However, Rayleigh scattering signals are formed by the superposition of light reflected by scatterers at different locations along the optical fiber with different time delays. Even within a single pulse duration, the positional information carried by the sampling points is not completely aliased, but rather contains subtle differences. The ResUNet++ deconvolutional neural network model used in this application learns these subtle differences in a large amount of data to fit the mapping relationship from low-resolution signals to high-resolution signals, thereby breaking through the aforementioned physical resolution limit at the algorithm level.

[0035] Step S40: The super-resolution distributed acoustic sensing signal is sequentially subjected to amplitude and phase calibration, vibration event spatial positioning optimization and visualization processing, and the location, quantity, waveform and range information of vibration events along the optical fiber are output.

[0036] Specifically, amplitude and phase calibration corrects the network's output signal to eliminate amplitude and phase deviations that may be introduced during network inference, ensuring the accuracy of the output signal. Calibration can be achieved by comparing the output signal with a known reference signal, calculating correction coefficients, and applying them to all output signals.

[0037] Vibration event spatial localization optimization involves determining the specific location of each vibration event on the optical fiber. In one embodiment, this optimization employs a strategy combining peak detection and threshold filtering. Specifically, the amplitude of each location in the super-resolution signal is first scanned along the fiber direction (i.e., the spatial dimension) to identify all local peaks. Then, an amplitude threshold is set, and peaks with amplitudes exceeding the threshold are identified as the locations of vibration events, while peaks below the threshold are identified as noise and discarded. This threshold can be dynamically set based on the overall noise level of the signal or a fixed value can be used.

[0038] Visualization processing presents the processed signal in an intuitive graphical format. In one embodiment, visualization processing includes generating waterfall plots and vibration localization curves. A waterfall plot is a graph displaying the signal intensity or phase distribution on a two-dimensional time-space plane. The horizontal axis represents the fiber location, the vertical axis represents time, and color or grayscale indicates the signal amplitude or phase value. This graph visually reflects the evolution of a vibration event over time. A vibration localization curve is a curve showing the distribution of vibration signal amplitude or energy along the fiber. The horizontal axis represents the fiber location, and the vertical axis represents the vibration amplitude or energy at that location. The area where the curve shows a peak is the vibration occurrence area.

[0039] After the above processing, the location, number, waveform and range information of vibration events along the optical fiber are finally output, thus completing the complete closed loop from raw signal acquisition to vibration event detection.

[0040] In some embodiments, the upsampling process adjusts the signal spatial sampling interval to a preset target value; the detrending process is used to eliminate the DC component and trend term in the signal; and the spatiotemporal slicing is used to convert the two-dimensional spatiotemporal signal into a signal segment of fixed size.

[0041] In one embodiment, the preset target value for upsampling is 0.125m, which means adjusting the spatial sampling interval of the signal to 0.125m through interpolation or other methods. This value is consistent with the spatial sampling interval of the input data during network training, ensuring that the actual acquired signal and the training data have the same sampling density in the spatial dimension, thereby guaranteeing that the network can correctly perform forward inference. Detrending processing can be implemented using methods such as multinomial fitting or high-pass filtering. Its core purpose is to subtract the fitted trend component from the signal, causing the signal to fluctuate around the zero baseline. The specific implementation of spatiotemporal slicing is as follows: the two-dimensional signal (one dimension is space, and one dimension is time) acquired during the entire measurement period is segmented according to a fixed-length time window. Each segment is treated as an independent sample, and the spatial dimension of each sample is kept intact so that the network can perform convolution operations in the spatial dimension.

[0042] In some embodiments, the visualization process includes generating a waterfall plot and a vibration localization curve. The waterfall plot is used to display the intensity or phase distribution of the super-resolution distributed acoustic sensing signal in a two-dimensional time-space plane, and the vibration localization curve is used to display the distribution curve of the vibration signal amplitude or energy along the optical fiber. The vibration event spatial localization optimization adopts a strategy combining peak detection and threshold screening. Specifically, an amplitude threshold is set, and the peak positions of the super-resolution distributed acoustic sensing signal with amplitudes exceeding the threshold are determined as the locations where vibration events occur, while noise signals below the threshold are removed.

[0043] In one embodiment, the waterfall plot is presented graphically, with the horizontal axis representing the distance along the optical fiber and the vertical axis representing time. The grayscale or color of a pixel represents the signal strength or phase value at that spatiotemporal location. On the waterfall plot, vibration events typically appear as stripes or bands that continuously appear along the time axis, and their horizontal positions correspond to the actual locations of the vibration events on the optical fiber. The vibration localization curve plots the signal amplitude or energy at various spatial locations within a certain moment or time window, and the peak position of the curve represents the possible location of the vibration event. The specific implementation of peak detection and threshold screening is as follows: first, all local maxima are found on the vibration localization curve; then, the amplitude of each maxima is checked one by one to see if it exceeds a preset threshold. Those exceeding the threshold are retained as valid vibration events, while those below are discarded as noise. The threshold can be set as a multiple of the root mean square of all signals, or an adaptive threshold algorithm can be used to dynamically calculate it based on the signal noise level.

[0044] In some embodiments, the ResUNet++ deconvolutional neural network model sequentially includes a Stem module, a ResNet module, a dilated spatial pyramid pooling module, a decoding module, and an output module. The Stem module is used to perform initial convolutional feature extraction on the input tensor. The ResNet module is used to extract deep signal features through residual connections and attention mechanisms. The dilated spatial pyramid pooling module is used to capture multi-scale contextual information through parallel convolutional layers with different dilation rates. The decoding module is used to restore the signal spatial resolution stepwise through upsampling and attention mechanisms. The output module is used to aggregate multi-scale features and output a super-resolution distributed acoustic sensing signal.

[0045] In one embodiment, the Stem module consists of two convolutional layers: the first layer has a 7×7 kernel size and a stride of 2, and the second layer has a 3×3 kernel size and a stride of 1. Each layer is followed by a batch normalization layer and a ReLU activation function. The output of the Stem module is fed into the ResNet module. The ResNet module contains four layers, each consisting of several residual blocks with 64, 128, 256, and 512 channels respectively. Each residual block contains two convolutional layers and a skip connection, which directly adds the input of the residual block to the output to alleviate the vanishing gradient problem. A channel attention module is also introduced in each residual block. This module calculates the weight coefficients of each channel through global average pooling and two fully connected layers, weighting each channel of the feature map so that the network focuses on important feature channels. The ASPP module contains four parallel convolutional layers with dilation rates of 1, 6, 12, and 18. The outputs of each convolutional layer are concatenated along the channel dimension, and then compressed to 256 channels using a 1×1 convolutional layer. The decoding module consists of four decoding layers, symmetrical to the four encoding layers of the ResNet module. Each decoding layer first upsamples the feature map by a factor of 2, then performs a skip connection with the feature map of the corresponding encoding layer, followed by two convolutional layers for feature refinement. The output module contains a 3×3 convolutional layer for feature aggregation and a 1×1 convolutional layer for mapping the feature map to a single-channel output signal.

[0046] In some embodiments, the low-resolution original signal is a backscattered Rayleigh light signal collected by the distributed acoustic sensing system after emitting a laser pulse into the sensing fiber with a predetermined pulse width, and obtained after interferometric detection and phase demodulation; the spatial resolution of the super-resolution distributed acoustic sensing signal is less than 2m, and it can distinguish multiple adjacent vibration events.

[0047] In one embodiment, the predetermined pulse width is 100 ns, corresponding to a physical resolution limit of 10 m. The spatial sampling interval is 6.25 ns, corresponding to a spatial sampling interval of approximately 0.625 m (calculated with an optical fiber refractive index of 1.5). Under this parameter configuration, the spatial resolution of the super-resolution signal processed by the method of this application can reach 1.5 m, capable of stably distinguishing two adjacent vibration events with a distance of not less than 1.5 m. In another embodiment, the predetermined pulse width is 37.5 ns, corresponding to a physical resolution limit of 3.75 m, and the spatial sampling interval is 25 ns, corresponding to a spatial sampling interval of approximately 2.5 m. A spatial resolution of less than 2 m can also be obtained after processing. The above numerical ranges are merely illustrative examples; the method of this application is not limited to specific pulse widths or sampling intervals, and its applicability depends on the data parameters used during network training.

[0048] In some embodiments, the pre-trained ResUNet++ deconvolutional neural network model is obtained through the following methods: Construct a ResUNet++ deconvolutional neural network model; Obtain a training dataset, which contains multiple sample pairs consisting of low-resolution distributed acoustic sensing signals and corresponding high-resolution label signals; The ResUNet++ deconvolutional neural network model is trained using the low-resolution distributed acoustic sensing signal as input and the corresponding high-resolution label signal as output. The loss function during training includes a data fidelity term and a physical constraint term. The physical constraint term is constructed based on the convolutional degradation model of the distributed acoustic sensing system, enabling the network to fit the inverse mapping of the fiber optic response function.

[0049] To illustrate the training process more clearly, a specific example is provided below. In one embodiment, a ResUNet++ deconvolutional neural network model with the above structure is constructed, and the network parameters are randomly initialized. The training dataset is generated using system simulation, obtaining a large number of training samples with accurate labels by simulating the response of the DAS system under different conditions. The simulation process is based on the aforementioned convolutional degradation model. A low-resolution signal is generated, where the width of the probe pulse p(t) is set to 100 ns, the fiber response function h(t) is generated based on the preset vibration event location, waveform, and intensity, and the noise n(t) is simulated using Gaussian white noise. The corresponding high-resolution tag signal is generated using a narrower pulse width (e.g., 10 ns), which has better spatial resolution than the low-resolution signal. The simulation scenarios cover different application scenarios with 0 to 4 vibration events. The waveforms (sine waves, square waves, pulse waves, etc.), locations (randomly distributed along the fiber), and lengths (random variations in vibration duration) of each vibration event are randomized to ensure sufficient diversity in the training data and improve the network's generalization ability. The number of training samples is no less than 5000 pairs. During training, the low-resolution signal is input into the network, the network output is obtained, the difference between the network output and the high-resolution tag is calculated, and the network parameters are updated using the backpropagation algorithm. The loss function during training is L = Ldata + λLphysics, where Ldata is the data fidelity term, expressed as mean squared error, used to measure the difference between the network output and the label signal; Lphysics is the physical constraint term, used to measure the difference between the network output after convolutional degradation and the network input; λ is a weight coefficient used to adjust the proportion of the physical constraint term in the total loss. The role of the data fidelity term is to make the network output approximate the label signal as closely as possible, and the role of the physical constraint term is to ensure that the network output satisfies the physical laws of DAS signal propagation, that is, the output signal should be consistent with the input signal after the same convolutional degradation process. This design allows the network to fit the inverse mapping of the fiber optic response function, rather than a simple end-to-end mapping, thus ensuring the physical reliability of the output signal. After iterative training until the loss function converges, the network weight parameters are saved, thus obtaining the pre-trained model.

[0050] In some embodiments, in the training dataset, the low-resolution distributed acoustic sensing signal of each sample pair corresponds to a first spatial resolution, and the high-resolution tag signal corresponds to a second spatial resolution, wherein the second spatial resolution is higher than the first spatial resolution; the training dataset is generated through system simulation and covers multiple application scenarios with different numbers of vibration events, and the waveform, position, and length of each vibration event are randomized; the convolutional degradation model is used to simulate the signal degradation process caused by the superposition of noise after the convolution of the probe pulse time profile and the fiber response function.

[0051] In one embodiment, the first spatial resolution is 10m, and the second spatial resolution is 1m. That is, the low-resolution input signal corresponds to a 10m resolution, and the high-resolution tag corresponds to a 1m resolution; the network needs to learn the mapping relationship from 10m resolution to 1m resolution. The simulation generation method is as follows: first, the scattering coefficient distribution and vibration event parameters (position, waveform, length) at various locations along the fiber are randomly generated; then, the received scattered signals under different pulse widths are calculated based on the convolutional degradation model. The signal calculated with a 100ns pulse width is used as the low-resolution input, and the signal calculated with a 10ns pulse width is used as the high-resolution tag. Since the physical resolution corresponding to a 10ns pulse width is higher than 100ns, the tag signal has a higher spatial resolution. During the simulation, the waveform type of each vibration event is randomly selected from sine waves, square waves, triangular waves, pulse waves, etc., and the positions are randomly distributed along the fiber (maintaining a certain distance between different events to avoid overlap), and the vibration length (duration) varies randomly within a preset range. Through the randomization process described above, the generated training dataset can cover a variety of practical application scenarios, enabling the network to learn signal characteristics under different vibration modes and distributions during training, thus exhibiting good generalization performance in practical applications. The convolutional degradation model is used in the simulation process to model the convolution relationship between the probe pulse and the fiber response function, i.e., to simulate the signal blurring effect caused by the finite width of the optical pulse propagating in the fiber.

[0052] This application also provides a distributed acoustic sensor signal reconstruction system, including: The signal acquisition module is used to acquire low-resolution raw signals from the distributed acoustic sensing system. The preprocessing module is used to perform normalization, detrending, spatiotemporal slicing and upsampling on the low-resolution raw signal in sequence, and convert it into tensor data. The super-resolution reconstruction module is loaded with a pre-trained ResUNet++ deconvolutional neural network model, which is used to perform forward inference on the tensor data and output a super-resolution distributed acoustic sensing signal; the pre-trained model is used to map the low-resolution distributed acoustic sensing signal into a super-resolution distributed acoustic sensing signal. The post-processing output module is used to sequentially perform amplitude and phase calibration, vibration event spatial positioning optimization, and visualization processing on the super-resolution distributed acoustic sensing signal, and output the location, quantity, waveform, and range information of vibration events along the optical fiber.

[0053] In one embodiment, the signal acquisition module specifically includes a narrow-pulse laser source, an interferometric detection module, and a phase demodulation module. These three sub-modules can be integrated on the same circuit board or are independently connected devices. The preprocessing module, super-resolution reconstruction module, and post-processing output module can be integrated into a digital signal processor or a graphics processor to achieve real-time signal processing. Data transmission between the above modules can be achieved through a data bus or through common interfaces such as Ethernet or USB. In one embodiment, the signal acquisition module transmits the acquired low-resolution raw signal to the processor where the preprocessing module is located via a high-speed data interface (such as PCIe). After preprocessing, the data is stored in the processor's memory. The super-resolution reconstruction module reads the preprocessed tensor data from memory for forward inference. After inference, the result is output to the post-processing output module for final processing. The entire process can achieve pipelined parallel processing, that is, the four stages of signal acquisition, preprocessing, super-resolution reconstruction, and post-processing are performed simultaneously to reduce overall processing latency and meet the application requirements of real-time monitoring.

[0054] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the distributed acoustic sensor signal reconstruction method as described in this application.

[0055] In one embodiment, the computer-readable storage medium may be a hard disk, solid-state drive, optical disk, USB flash drive, or SD card, etc. When the computer program in the storage medium is executed by the processor, it first receives the low-resolution raw signal from the signal acquisition module, and then sequentially performs preprocessing, super-resolution reconstruction, and post-processing steps, finally outputting the location, quantity, waveform, and range information of the vibration event. This program can run on the embedded processor of the DAS system or on a general-purpose computer; this application does not impose any limitations on this.

[0056] This application also provides a computer program product, including a computer program or computer executable instructions, which, when executed by a processor, implement the distributed acoustic sensor signal reconstruction method as described in this application.

[0057] In one embodiment, the computer program product is provided as a software installation package, which can be downloaded by the user via a network or obtained from physical media. The installation package contains program code implementing the above-described method and related configuration information. After the user installs the program product on the data processing terminal of the DAS system, the terminal will have the function of distributed acoustic sensor signal hyperspatial resolution reconstruction.

[0058] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical applications, or technological improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for reconstructing distributed acoustic sensor signals, characterized in that, include: Acquire low-resolution raw signals from a distributed acoustic sensing system; The low-resolution original signal is sequentially normalized, detrended, spatiotemporally sliced, and upsampled to be converted into tensor data. The tensor data is input into a pre-trained ResUNet++ deconvolutional neural network model for forward inference, and outputs a super-resolution distributed acoustic sensing signal; the pre-trained model is used to map the low-resolution distributed acoustic sensing signal into a super-resolution distributed acoustic sensing signal. The super-resolution distributed acoustic sensing signal is sequentially subjected to amplitude and phase calibration, vibration event spatial positioning optimization, and visualization processing to output the location, quantity, waveform, and range information of vibration events along the optical fiber.

2. The method according to claim 1, characterized in that: The upsampling process adjusts the signal spatial sampling interval to a preset target value; the detrending process is used to eliminate the DC component and trend term in the signal; the spatiotemporal slicing is used to convert the two-dimensional spatiotemporal signal into a signal segment of fixed size.

3. The method according to claim 1, characterized in that: The visualization process includes generating waterfall plots and vibration localization curves. The waterfall plots are used to display the intensity or phase distribution of the super-resolution distributed acoustic sensing signal in a two-dimensional time-space plane, and the vibration localization curves are used to display the distribution curve of the vibration signal amplitude or energy along the optical fiber. The vibration event spatial localization optimization adopts a strategy combining peak detection and threshold filtering. Specifically, an amplitude threshold is set, and the peak positions of the super-resolution distributed acoustic sensing signal with amplitudes exceeding the threshold are determined as the locations of vibration events, while noise signals below the threshold are eliminated.

4. The method according to claim 1, characterized in that: The ResUNet++ deconvolutional neural network model sequentially includes a Stem module, a ResNet module, a dilated spatial pyramid pooling module, a decoding module, and an output module. The Stem module is used to perform initial convolutional feature extraction on the input tensor. The ResNet module is used to extract deep signal features through residual connections and attention mechanisms. The dilated spatial pyramid pooling module is used to capture multi-scale contextual information through parallel convolutional layers with different dilation rates. The decoding module is used to restore the signal spatial resolution step by step through upsampling and attention mechanisms. The output module is used to aggregate multi-scale features and output a super-resolution distributed acoustic sensing signal.

5. The method according to claim 1, characterized in that: The low-resolution original signal is the backscattered Rayleigh light signal collected by the distributed acoustic sensing system after emitting a laser pulse into the sensing fiber with a predetermined pulse width, and obtained after interferometric detection and phase demodulation; the super-resolution distributed acoustic sensing signal has a spatial resolution of less than 2m and can distinguish multiple adjacent vibration events.

6. The method according to claim 1, characterized in that: The pre-trained ResUNet++ deconvolutional neural network model was obtained in the following way: Construct a ResUNet++ deconvolutional neural network model; Obtain a training dataset, which contains multiple sample pairs consisting of low-resolution distributed acoustic sensing signals and corresponding high-resolution label signals; The ResUNet++ deconvolutional neural network model is trained using the low-resolution distributed acoustic sensing signal as input and the corresponding high-resolution label signal as output. The loss function during training includes a data fidelity term and a physical constraint term. The physical constraint term is constructed based on the convolutional degradation model of the distributed acoustic sensing system, enabling the network to fit the inverse mapping of the fiber optic response function.

7. The method according to claim 6, characterized in that: In the training dataset, the low-resolution distributed acoustic sensing signal of each sample pair corresponds to the first spatial resolution, and the high-resolution label signal corresponds to the second spatial resolution, which is higher than the first spatial resolution. The training dataset is generated through system simulation and covers multiple application scenarios with different numbers of vibration events. The waveform, position, and length of each vibration event are randomized. The convolutional degradation model is used to simulate the signal degradation process caused by the superposition of noise after the convolution of the probe pulse time profile and the fiber response function.

8. A distributed acoustic sensor signal reconstruction system, characterized in that, include: The signal acquisition module is used to acquire low-resolution raw signals from the distributed acoustic sensing system. The preprocessing module is used to perform normalization, detrending, spatiotemporal slicing and upsampling on the low-resolution raw signal in sequence, and convert it into tensor data. The super-resolution reconstruction module is loaded with a pre-trained ResUNet++ deconvolutional neural network model, which is used to perform forward inference on the tensor data and output a super-resolution distributed acoustic sensing signal; the pre-trained model is used to map the low-resolution distributed acoustic sensing signal into a super-resolution distributed acoustic sensing signal. The post-processing output module is used to sequentially perform amplitude and phase calibration, vibration event spatial positioning optimization, and visualization processing on the super-resolution distributed acoustic sensing signal, and output the location, quantity, waveform, and range information of vibration events along the optical fiber.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the distributed acoustic sensor signal reconstruction method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, It includes a computer program or computer-executable instructions, which, when executed by a processor, implement the distributed acoustic sensing signal reconstruction method as described in any one of claims 1 to 7.

Images