A high-resolution distributed acoustic vibration demodulation method based on frequency-sweep light depth anti-fading

Abstract

This invention discloses a high-resolution distributed acoustic-vibration demodulation method based on sweep-frequency optical deep anti-fading. First, continuous linear sweep-frequency probe light is injected into an optical fiber, and the original backscattered Rayleigh signal containing information distributed along the fiber is acquired through coherent beat frequency acquisition. Then, the original backscattered Rayleigh signal undergoes frequency domain demultiplexing and matched-filter pulse compression processing to reconstruct multiple high-resolution time-domain complex envelope signals from a single sweep. Next, the high-resolution time-domain complex envelope signals are input into a deep anti-fading demodulation network, where amplitude-phase joint modeling achieves adaptive suppression of random fading and noise in the time-domain complex envelope signal, outputting a de-fading phase sequence. Finally, the de-fading phase sequence is used for phase difference demodulation, and combined with an optical time-of-flight positioning model, to achieve highly robust demodulation and spatial positioning of vibration events. This method maintains high resolution and robustness even under strong noise, long distance, and deep fading conditions, making it suitable for distributed acoustic-vibration and ultrasonic monitoring scenarios.

Description

Technical Field

[0001] This invention belongs to the fields of fiber optic sensing, Rayleigh scattering signal processing, and artificial intelligence digital signal processing technology, and relates to a high-resolution distributed acoustic-vibration demodulation method based on sweep frequency optical depth anti-fading. Background Technology

[0002] Rayleigh scattering distributed optical fiber sensing (DOFS), a technology that utilizes optical fibers to achieve high-sensitivity vibration / strain monitoring along a fiber optic line, has been widely used in structural health monitoring, earthquake detection, industrial equipment monitoring, perimeter security, and other fields. Existing systems mainly include those based on phase demodulation. COTD based on frequency demodulation, and COTD using chirped pulses. Solutions such as [list of solutions]. With the increasing demands of industrial monitoring, systems need to simultaneously meet multiple requirements, including long distance, high resolution, high sensitivity, and high dynamic range. However, traditional frequency-scanning COTDRs suffer from long scan cycles and insufficient real-time performance; while pulse-type [solutions / solutions]... It also faces the inherent contradiction between pulse width, spatial resolution and detection energy.

[0003] Regardless of whether phase demodulation or frequency demodulation is used, Rayleigh scattering-based DOFS systems are inevitably affected by interference fading. Rayleigh scattering signals are formed by the random coherent superposition of numerous micro-scattering centers, and their amplitude exhibits strong random fluctuations in space. Deep fading occurs at some locations, leading to phase demodulation distortion or signal loss, creating detection blind zones along the line. To address this problem, existing research has proposed methods such as multi-frequency detection, multi-sub-pulse, or frequency diffusing synthesis, introducing redundant information in the frequency or time dimensions to mitigate the fading effect. However, these methods typically require multiple transmissions or multi-sub-band synchronous acquisition, resulting in high system complexity, and often rely on fixed weights or empirical rules for signal synthesis. When the fading location or noise distribution changes over time, the synthesis effect is difficult to adaptively adjust, and it remains challenging to stably suppress random fading under complex operating conditions.

[0004] In recent years, deep learning has gradually demonstrated its potential in the field of distributed fiber optic sensing, and has been applied to tasks such as event recognition, seismic wave classification, spectral domain denoising, decorrelation, and data compression. Among these, convolutional neural networks (CNNs), autoencoders (AEs), Transformer structures, and weakly supervised denoising networks have shown outstanding performance in processing spatiotemporally structured DAS data. However, existing research focuses primarily on backend data interpretation and noise suppression, such as improving event classification accuracy, extracting high-dimensional acoustic signature features, or reducing background noise. Few studies have directly embedded deep learning into the front-end physical demodulation chain to replace traditional manually designed frequency band weighting, vector synthesis, or phase compensation strategies. Especially in the multi-band Rayleigh scattering signal fusion stage, traditional methods such as empirical weights, energy ratios, SNR estimation, or fixed filter coefficients are still commonly used. These methods cannot adaptively model fading statistics based on different spatial and temporal locations, nor can they intelligently adjust amplitude and phase information simultaneously, thus making it difficult to maintain robust demodulation performance in high-noise, high-fading environments.

[0005] Therefore, while taking into account high resolution, high bandwidth and long-distance sensing capabilities, how to overcome the limitations of traditional multi-frequency and multi-pulse synthesis methods and construct an acoustic-vibration demodulation method with adaptive anti-fading capabilities remains a key issue that urgently needs to be addressed. Summary of the Invention

[0006] To address the aforementioned technical problems in existing technologies, this invention proposes a high-resolution distributed acoustic-vibration demodulation method based on swept-frequency optical depth anti-fading. Broadband Rayleigh scattering information is obtained through linearly frequency-modulated chirped signals, and signal compression and subband decomposition are achieved using a matched filter compression module. Furthermore, a deep learning network is combined to perform weighted regression and phase compensation on each subband, achieving a three-in-one approach: decoupling of sensing distance and detection bandwidth, improvement of spatial resolution and signal-to-noise ratio, and optimization of system anti-interference capabilities. The specific technical solution is as follows:

[0007] A high-resolution distributed acoustic-vibration demodulation method based on swept-frequency optical depth anti-fading includes the following steps:

[0008] Step 1: Inject continuous linear sweep frequency probe light into the optical fiber and obtain the original back Rayleigh scattering signal containing information distributed along the optical fiber through coherent beat frequency;

[0009] Step 2: Perform frequency domain demultiplexing and matched filtering pulse compression on the original backscattered Rayleigh signal to reconstruct the single-sweep signal into multiple high-resolution time-domain complex envelope signals;

[0010] Step 3: Input the high-resolution time-domain complex envelope signal into the deep anti-fading demodulation network, and achieve adaptive suppression of random fading and noise of the time-domain complex envelope signal through amplitude and phase joint modeling, and output the defading phase sequence;

[0011] Step 4: Demodulate the phase difference using the defading phase sequence and combine it with the optical time-of-flight positioning model to achieve demodulation and spatial positioning of vibration events.

[0012] Furthermore, step 1 specifically includes:

[0013] Step 1.1: A narrow linewidth laser is used as the light source to output continuous light. The linear frequency modulated electrical signal generated by the arbitrary waveform generator is input into the electro-optic modulator to modulate the frequency of the continuous light, so that the laser frequency scans linearly with time and oscillates between the preset start frequency and end frequency, so that each instantaneous frequency point corresponds to a unique equivalent optical path.

[0014] Step 1.2: The modulated sweep probe light is split by an optical fiber coupler and injected into the optical fiber under test. The probe light propagates unidirectionally along the optical fiber and undergoes Rayleigh scattering with randomly distributed scattering centers in the fiber core. Under the action of ultrasonic events or vibration disturbances, phase modulation related to local strain changes is generated. Among them, the back Rayleigh scattering component propagating in the opposite direction of incident carries the scattering intensity distribution and phase disturbance information at each position along the line, and is coupled back to the detection optical path at the optical fiber port.

[0015] Step 1.3: Another part of the continuous light output from the laser is used as the local oscillator light and directly introduced into the balanced detector through an independent optical path to interfere with the back Rayleigh scattered light from the fiber under test. Since the probe light is linearly swept, a time-varying difference frequency signal is formed between the local oscillator light and the back Rayleigh scattered light. The balanced detector performs differential detection on the intensity of the two light paths and outputs an electric domain beat frequency interference signal with amplitude representing the scattering intensity and phase representing the fiber strain disturbance.

[0016] Step 1.4: The electric domain beat frequency interference signal output by the balanced detector is pre-amplified and anti-aliasing filtered before being sent to the data acquisition card. Synchronous sampling is performed at a sampling rate of not less than twice the upper limit of the beat frequency. Under the multi-channel architecture, the beat frequency signals of multiple probe lights or multiple optical fibers are acquired in parallel to obtain the time-continuous original back Rayleigh scattering signal.

[0017] Furthermore, step 2 specifically includes:

[0018] Step 2.1: Demultiplex the original backscattered Rayleigh signal in the frequency domain according to the frequency sweep rule: First, based on the different effective length segments in the optical fiber corresponding to different frequency intervals, the original backscattered Rayleigh signal represented in the frequency domain is divided into several sub-frequency modulated pulse frequency domain signals according to the frequency interval using this mapping. Then, a fast inverse Fourier transform is performed on each sub-frequency modulated pulse frequency domain signal to obtain the time domain signal of the sub-frequency modulated pulse.

[0019] Step 2.2: Based on linear frequency modulation matched filtering technology, the preset sub-pulse signal is deconvoluted and conjugated, and then convolved with the time domain signal of each sub-frequency modulation pulse to perform pulse compression, and output a high-resolution time domain complex envelope signal sequence after compression.

[0020] Furthermore, step 3 specifically includes:

[0021] Step 3.1: Set the extent of the local data block within the 3D window;

[0022] Step 3.2: Divide the multi-channel high-resolution time-domain complex envelope signal into several local data blocks according to the spatial-temporal dimension, and then input them into the deep learning anti-fading demodulation network to complete feature initialization, coarse feature extraction, fine feature reconstruction and complex domain amplitude and phase recovery in sequence, and train to output the defading phase sequence.

[0023] Furthermore, in step 3.2, a low-level feature description of the local data block is established through shallow 3D convolution to complete feature initialization.

[0024] Furthermore, in step 3.2, multi-scale dilated convolution is used to capture the fading structure and weak vibration texture of the initial features, thus completing the coarse feature extraction.

[0025] Furthermore, in step 3.2, the details of the deep fading region in the coarse features are repaired by using a three-dimensional encoding-decoding structure in conjunction with skip connections, and the defading complex envelope is obtained by linear convolution mapping to complete the fine feature reconstruction. The three-dimensional encoding-decoding structure consists of multi-layer two-dimensional convolution operations, nonlinear activation operations and feature normalization processing.

[0026] Furthermore, in step 3.2, the defading complex envelope and the real high signal-to-noise ratio complex envelope are used together for loss construction. The amplitude reconstruction error and phase consistency error are combined to perform reverse optimization on the deep learning anti-fading demodulation network. Finally, the defading complex envelope and defading phase sequence of each local data block are seamlessly spliced ​​along the spatial and temporal dimensions to generate the final three-dimensional defading complex envelope covering the entire fiber length and the corresponding continuous phase sequence.

[0027] Furthermore, step 4 specifically includes:

[0028] Step 4.1: By performing phase difference demodulation on adjacent spatial sampling points of the defading phase sequence, the common phase component is suppressed and the local disturbance response is highlighted, and the vibration response at each position along the optical fiber is extracted.

[0029] Step 4.2: Perform energy statistics on the vibration response within the time sliding window, construct event judgment features, and when the energy exceeds the preset threshold, determine that there is a vibration event in the corresponding spatial location and time window, and then output the spatiotemporal location set of candidate events and their corresponding energy features.

[0030] Step 4.3: Combining the optical time of flight and optical path mapping relationship, map the time index corresponding to the event to the actual optical fiber spatial location, and then determine the precise location of the event through the energy peak;

[0031] Step 4.4: Perform time-frequency analysis or statistical feature extraction on the local vibration response after positioning, and output the event type and its location.

[0032] The advantages and beneficial effects of this invention are as follows:

[0033] 1. This invention proposes a single-shot large-bandwidth detection mechanism based on continuous linear frequency sweep light. An equivalent multi-frequency scanning effect can be mathematically synthesized through a single frequency sweep, which effectively improves the time duty cycle and spectrum utilization efficiency of the detection light. It breaks through the strong coupling relationship between detection distance, system bandwidth and real-time performance in traditional multi-frequency and multi-pulse schemes, and is suitable for long-distance, high-bandwidth distributed acoustic and vibration monitoring scenarios.

[0034] 2. This invention combines frequency domain demultiplexing and linear frequency modulation matched filtering technology to achieve high-resolution pulse compression of sub-frequency modulated pulses, so that the spatial resolution is determined by the sweep bandwidth rather than the pulse width. Without reducing the detection energy, it significantly improves the spatial resolution and signal-to-noise ratio, overcoming the inherent contradiction that resolution and signal-to-noise ratio are difficult to balance in traditional pulse-type distributed sensing systems.

[0035] 3. This invention is the first to directly introduce deep learning into the front-end physical demodulation link of Rayleigh scattering signals. By performing spatial and temporal block processing on multi-channel complex envelope signals and using multi-scale convolution and encoder-decoder network structures to adaptively model random fading modes and noise distribution, it achieves intelligent repair of deep fading regions and avoids the lack of robustness caused by traditional multi-frequency synthesis methods that rely on fixed weights or empirical rules.

[0036] 4. This invention employs a three-dimensional data block partitioning and a parallel inference strategy without synchronization dependencies, enabling the deep anti-fading demodulation network to fully utilize the parallel computing capabilities of CPU / GPU. While ensuring demodulation accuracy, it significantly improves system throughput and real-time processing performance, enhancing the scalability and practicality of the method in engineering applications.

[0037] 5. Based on the deep anti-fading demodulation results, this invention combines phase difference demodulation with an optical time-of-flight positioning model to achieve highly reliable demodulation and accurate positioning of ultrasonic disturbances, mechanical vibrations, and intrusion events. Thanks to the significant improvement in signal-to-noise ratio and phase continuity brought about by front-end anti-fading processing, this method can maintain good stability and positioning accuracy even under strong noise, long distance, and deep fading conditions. Attached Figure Description

[0038] Figure 1 This is a flowchart of a high-resolution distributed acoustic-vibration demodulation method based on swept-frequency optical depth anti-fading according to an embodiment of the present invention.

[0039] Figure 2 This is a schematic diagram of the structure of light source emission and coherent detection in an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram of the high-resolution signal reconstruction module according to an embodiment of the present invention.

[0041] Figure 4 This is a schematic diagram of the structure of multi-channel complex envelope signal partitioning according to an embodiment of the present invention.

[0042] Figure 5 This is a schematic diagram of the structure of a deep learning anti-fading demodulation network according to an embodiment of the present invention.

[0043] Figure 6 This is a schematic diagram of the coarse feature extraction module according to an embodiment of the present invention.

[0044] Figure 7 This is a schematic diagram of the structure of the fine feature extraction module in an embodiment of the present invention.

[0045] Figure 8 This is a schematic diagram of the signal demodulation and event localization process according to an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0047] like Figure 1 As shown in this embodiment, a high-resolution distributed acoustic-vibration demodulation method based on swept-frequency optical depth anti-fading aims to achieve a three-in-one approach: decoupling of sensing distance and detection bandwidth, improvement of spatial resolution and signal-to-noise ratio, and optimization of system anti-interference capability. The method includes the following steps:

[0048] Step 1: As Figure 2 As shown, a continuously sweeping probe light is injected into the fiber under test. By controlling the laser frequency to scan linearly with time, each frequency point corresponds to a fixed optical path. After the probe light propagates along the fiber, its backscattered Rayleigh signal (RS signal) and the local oscillator light coherently beat in a balanced detector, generating a raw interference signal containing phase perturbation and scattering intensity distribution. Finally, multi-channel beat data is synchronously acquired through a data acquisition card, and the raw backscattered Rayleigh signal that can be used for subsequent demultiplexing and demodulation is output.

[0049] Step 1 specifically includes:

[0050] Step 1.1: A narrow-linewidth laser is used as the light source to output continuous light with a fixed center wavelength and narrow spectral linewidth. A linearly frequency-modulated electrical signal generated by an arbitrary waveform generator is input into an electro-optic modulator to frequency-modulate the continuous light, causing the laser frequency to scan linearly over time. This scanning cyclically between a preset start frequency and a stop frequency ensures that each instantaneous frequency point corresponds to a unique equivalent optical path, achieving the transmission of a single, large-bandwidth continuous frequency-scanning probe beam. This allows for the mathematical synthesis of a multi-frequency scanning effect. The specific implementation is as follows:

[0051] Sending a single, high-bandwidth linear frequency modulated electrical signal, covering the frequency range of all sub-pulses, can be expressed by the following formula:

[0052] ,

[0053] in, It is the complex envelope signal of a linear frequency modulated pulse; Indicates the linear frequency modulated pulse at Instantaneous frequency at time; The linear frequency modulation slope is defined as follows: , For frequency sweep, The pulse duration.

[0054] For the above formula Perform a Fourier transform to obtain the frequency domain. .

[0055] Step 1.2: The modulated swept probe light is split by an optical fiber coupler and injected into the fiber under test. The probe light propagates unidirectionally along the fiber, undergoing Rayleigh scattering with randomly distributed scattering centers within the fiber core. Under the influence of ultrasonic events or vibration disturbances, phase modulation related to local strain changes is generated. The backscattering component, propagating in the opposite direction of incident light, carries the scattering intensity distribution and phase disturbance information at various locations along the line and is coupled back to the detection optical path at the fiber port. The specific implementation is as follows:

[0056] The Rayleigh impulse response generated in the fiber optic test domain (FUT) can be expressed by the following formula:

[0057] ,

[0058] in, Indicates the Rayleigh scattering response of the FUT; The scattering amplitude at the i-th scattering point is a random variable. This indicates the phase introduced by the propagation distance; Indicates the round-trip time from the scattering point; This indicates the round-trip time for locating the scattering point.

[0059] The phase change generated under ultrasonic or vibrational action, which is related to local strain, can be expressed by the following formula:

[0060] ,

[0061] Where n represents the refractive index of light, with a typical value of about 1.45. When the optical fiber is stretched, the effective optical path changes due to the change in refractive index. This indicates the sensitivity to changes in refractive index caused by strain, with a typical value of approximately [value missing]. ; Indicates the wavelength of the incident light; Indicates local strain changes; This indicates the spatial sampling interval; the smaller the value, the higher the resolution of the accompanying fiber.

[0062] Step 1.3: Another portion of the continuous light output from the laser is used as the local oscillator light and directly introduced into the balanced detector through an independent optical path. It interferes with the backscattered Rayleigh light from the fiber under test within the detector arm. Since the probe light is linearly swept, a time-varying difference frequency signal is formed between the local oscillator light and the backscattered light. The balanced detector performs differential detection of the two light intensities, effectively suppressing common-mode noise and outputting an electrical domain beat frequency interference signal whose amplitude represents the scattering intensity and whose phase represents the fiber strain disturbance. The specific implementation is as follows:

[0063] The local oscillator light enters the optical fiber, and the probe light returns to generate a backward Rayleigh signal, as shown in the following formula:

[0064] ,

[0065] ,

[0066] in It is Ben Zhenguang; This indicates the amplitude of the local oscillator light used for coherent detection; Indicates the optical carrier frequency; The complex electric field represents the total Rayleigh scattered light returning from the optical fiber; Indicates the fiber attenuation coefficient; This indicates that the return pulse of the fiber insertion probe pulse is delayed. .

[0067] The local oscillator light and the back Rayleigh signal are mixed and square-law detected in a balanced detector to automatically generate a beat frequency signal. Since RS is too weak, its self-multiplication term is ignored. Subtracting the average value from the signal leaves only the time-varying component, yielding the electric domain beat frequency interference signal where amplitude represents scattering intensity and phase represents strain perturbation. The specific formula can be expressed as:

[0068] ,

[0069] ,

[0070] ,

[0071] in It is the response coefficient of the photodetector, which represents the efficiency of the photodetector in converting light signals into electrical signals. It is the frequency domain representation of the original RS signal.

[0072] Step 1.4: The beat frequency electrical signal output from the balanced detector is pre-amplified and anti-aliasing filtered before being sent to the data acquisition card. Synchronous sampling is performed at a sampling rate no less than twice the upper limit of the beat frequency. In a multi-channel architecture, beat frequency signals from multiple probe beams or multiple fiber segments are acquired in parallel to obtain a time-continuous raw Rayleigh signal. This raw Rayleigh signal serves as the input data basis for subsequent frequency domain multiplexing / demultiplexing processing and high-resolution demodulation. The specific formula can be expressed as:

[0073] ,

[0074] ,

[0075] Among them, subscript This represents the m-th sub-pulse; Let represent the frequency domain weighting function corresponding to the m-th sub-pulse.

[0076] Step 2: As Figure 3 As shown, the original backscattered Rayleigh signal is demultiplexed in the frequency domain by using a fast Fourier transform to obtain the scattering information at each spatial location. The sub-pulse time-domain Rayleigh signals are then reassembled according to the sub-pulse index to obtain multiple sub-pulse time-domain Rayleigh signals distributed along the optical fiber. Subsequently, based on linear frequency modulated matched filtering technology, pulse compression is performed on each sub-pulse signal. Through conjugate filtering and time defolding operations, a compression result with a narrow main lobe and high spatiotemporal resolution is obtained, outputting a high-resolution time-domain complex envelope signal sequence.

[0077] Step 2 specifically includes:

[0078] Step 2.1: Perform frequency domain demultiplexing based on the frequency sweep pattern. First, determine the different frequency ranges. This corresponds to different effective length segments in the optical fiber, and then this mapping is used to convert the frequency domain signal. The signal is divided into M sub-bands according to frequency intervals; then, an inverse Fourier transform is performed on each sub-band signal to obtain the time-domain signal of the sub-pulse. The specific formula can be expressed as:

[0079] ,

[0080] ,

[0081] Step 2.2: The impulse response of the matched filter is a response to the frequency domain sub-response. After performing deconvolution and conjugation, convolution is applied to each sub-pulse to achieve pulse compression. The width of the main lobe after pulse compression determines the spatial resolution, and the specific formula can be expressed as:

[0082]

[0083]

[0084]

[0085] in This represents the compressed time-domain signal; Represents the speed of light in a vacuum, usually taken as ; This represents the bandwidth of the m-th sub-pulse.

[0086] Step 2.3: Map the compressed signal to the spatial axes and fuse all sub-pulses to form the final complex envelope. Each scan cycle yields a high-resolution complex envelope along the fiber space. Stacking K consecutive frames of data produces a three-dimensional input tensor, which serves as the standard input for subsequent deep anti-fading networks. The specific formula can be expressed as:

[0087] ,

[0088] ,

[0089] ,

[0090] Where z represents the spatial position of the optical fiber; This means that the compression result of each sub-pulse is mapped to the spatial domain and fused to obtain a single frame of high-resolution complex envelope along the optical fiber. 'c' represents the three-dimensional input tensor obtained by stacking K consecutive frames of data; 'c' represents the channel dimension.

[0091] Step 3: Divide the pulse-compressed multi-channel time-domain complex envelope signal into several local data blocks according to the space-time dimension to achieve parallel inference without synchronization dependencies between data blocks, thereby maximizing CPU / GPU resource utilization and improving throughput. Figure 4As shown. Then, the input is fed into a deep learning anti-fading demodulation network. A multi-scale feature extraction structure is used to extract Rayleigh scattering amplitude and phase features. Through dilated convolution, encoder-decoder networks, and skip connections, fading modes, noise distribution, and local vibration perturbation patterns are modeled. During the training phase, an ideal complex envelope signal generated with a high signal-to-noise ratio is used as a reference label. Complex domain reconstruction loss and phase consistency loss are constructed to achieve adaptive recovery of deep fading, random phase perturbations, and strong noise regions. Finally, the output is a de-faded, high-fidelity, phase-continuous complex envelope signal and its corresponding continuous phase sequence, as shown. Figure 5 As shown.

[0092] Step 3 specifically includes:

[0093] Step 3.1: Divide the input tensor into local data blocks according to the 3D window. Divide each block into 128 sampling points in the z-direction; divide each block into 16 consecutive frames in the k-direction; and retain the entire block in the c-direction as the depth channel. Finally, output the set of data blocks. The mathematical expression is: These blocks are completely independent of each other and do not require global context fusion, thus enabling parallel distribution of computing tasks across CPUs / GPUs without synchronization dependencies.

[0094] Step 3.2: Process the pulse-compressed three-dimensional complex envelope signal The signal is divided into independent blocks based on spatial and temporal dimensions. These blocks are then input into a deep learning anti-fading demodulation network to sequentially perform feature initialization, coarse feature extraction, fine feature reconstruction, and complex domain amplitude-phase recovery. Feature initialization establishes a low-level feature description of the block through shallow 3D convolutions. Coarse feature extraction captures fading structures and weak vibrational textures using multi-scale dilated convolutions. Fine feature extraction repairs details in deep fading regions using a 3D encoder-decoder structure combined with skip connections. During training, complex domain amplitude-phase reconstruction loss and phase consistency loss are constructed to achieve robust recovery from noise and random phase perturbations, ultimately outputting a fading-free, high-fidelity, and phase-continuous 3D complex envelope signal. The specific steps are as follows:

[0095] Step 3.2.1: Initialize the original image features. First, each block is initialized individually during the inference phase. Then, shallow 3D convolutions are performed on the blocks as feature initialization. The specific formula is shown below:

[0096] ,

[0097] in, Indicates ReLU activation; kernel preserved. Resolution, Directional channel mixing; These are the initial features used to enter the coarse feature extraction module.

[0098] Step 3.2.2: Input the initial features into the coarse feature extraction module, such as... Figure 6 As shown, the input features contain dense local scattering information. The network abstracts this information step by step through multiple layers of convolution and pooling, mapping local patterns into more representative sparse features. This structure can expand the receptive field while preserving key spatial statistical structures, thereby achieving effective extraction of wide-scale fading patterns and obtaining coarse-layer features. This lays the feature foundation for subsequent fine feature reconstruction.

[0099] Step 3.2.3: Input the coarse features into the fine feature extraction module, such as... Figure 7 As shown, this module adopts an encoder-decoder structure. The fine feature extraction module consists of multi-layer two-dimensional convolution operations (conv2D), nonlinear activation operations (ReLU), and feature normalization (BN) processing, used to progressively abstract the local Rayleigh scattering structure in the input feature map. In the decoding stage, upsampling operations improve the spatial resolution of the feature map. Subsequently, linear convolution mapping transforms the deep features into the real and imaginary parts of the complex envelope, obtaining... and This forms the local complex envelope prediction result, as shown in the following formula:

[0100] ,

[0101] Step 3.2.4: Input the local complex envelope prediction results into the output module, and obtain the de-fading complex envelope and de-fading phase sequence through independent amplitude-phase decoupling structures; in this step, the predicted complex envelope and the true high SNR complex envelope are used together for loss construction, and the network is inversely optimized by combining amplitude reconstruction error and phase consistency error; finally, the de-fading complex envelopes of each block are... With defading phase sequence Seamlessly stitched along the spatial and temporal dimensions, a final three-dimensional de-fading complex envelope covering the entire fiber length is generated. and the corresponding continuous phase sequence This provides stable and high-fidelity feature input for subsequent differential phase demodulation, event recognition, and localization.

[0102] Step 4: As Figure 8As shown, based on a continuous phase sequence, phase difference is performed on adjacent sampling points in the spatial dimension to suppress common phase components, extracting local phase changes caused by external ultrasonic, mechanical vibration, or intrusion events, and obtaining a vibration response sequence distributed along the optical fiber. Event judgment features are constructed by performing energy statistics on the vibration response within a time sliding window. When the energy exceeds a preset threshold, a vibration event is determined to exist at the corresponding location. Subsequently, the time-of-flight index corresponding to the event is mapped to the actual optical fiber location based on the optical path-time mapping relationship, and the precise location of the event is determined by combining the energy distribution features. Finally, time-frequency analysis or statistical feature extraction is performed on the located local vibration response to achieve the identification and location of different disturbance event types.

[0103] Step 4 specifically includes:

[0104] Step 4.1: Phase difference demodulation to obtain the vibration response sequence along the line. This is achieved by analyzing the continuous phase sequence... Phase difference is performed on adjacent spatial sampling points to suppress common phase components and highlight local disturbance responses, thus obtaining vibration response sequences at various locations along the optical fiber. This vibration response directly characterizes the local phase change caused by external ultrasonic, mechanical vibration, or intrusion events, as shown in the following formula:

[0105] ,

[0106]

[0107] in, This represents the system spatial sampling interval.

[0108] Step 4.2: Event Detection and Spatiotemporal Feature Extraction. Energy statistics are performed on the vibration response within a sliding time window to construct event judgment features. When the energy exceeds a preset threshold, a vibration event is determined to exist within the corresponding spatial location and time window, and a set of candidate event spatiotemporal locations is output. And its corresponding energy characteristics. The specific formula is shown below:

[0109] ,

[0110] Step 4.3: Spatial Location Inversion Based on Optical Path Mapping. Based on the system's optical path-time mapping relationship, the time index corresponding to the event is mapped to the actual spatial location. Then, the final event location is determined through the energy peak value, and the estimated event spatial location is output. The specific formula is as follows:

[0111] ,

[0112] ,

[0113] in, The speed of light in a vacuum. The effective refractive index of the optical fiber. This is the flight time index for the corresponding event.

[0114] Step 4.4: Event Type Identification and Result Output. Time-frequency analysis or statistical feature extraction is performed on the local vibration response to distinguish between different types of events, such as ultrasonic disturbance, mechanical vibration, or intrusion. Combining the vibration spectrum distribution and energy characteristics, the event type and its location are output. The specific formula is as follows:

[0115] ,

[0116] In summary, the high-resolution distributed acoustic-vibration demodulation method based on continuous frequency sweeping light and deep anti-fading of the present invention includes four key steps:

[0117] First, a continuous linear frequency sweep probe light is injected into the fiber under test. By using a narrow linewidth laser, an electro-optic modulator and a coherent detection structure, the probe light covers a wide bandwidth frequency domain during a single frequency sweep. By coherently beating the local oscillator light with the back Rayleigh scattered light, the original Rayleigh scattering signal containing information about the spatial distribution along the fiber is obtained, thus achieving signal acquisition with high duty cycle and high spectral utilization.

[0118] Next, the original Rayleigh scattering signal is subjected to frequency domain multiplexing / demultiplexing. Through frequency domain grouping and linear frequency modulation matched filter pulse compression, the single continuous frequency sweep signal is mathematically reconstructed into multiple equivalent sub-frequency modulation pulse responses, and a high-resolution time-domain complex envelope signal whose spatial resolution is determined by the frequency sweep bandwidth is obtained. Thus, high-resolution signal reconstruction is achieved without reducing the detection energy.

[0119] Then, the high-resolution complex envelope signal is divided into multiple data blocks according to the spatial-temporal dimension and input into a deep learning anti-fading demodulation network. Through multi-scale feature extraction, encoding-decoding structure and amplitude-phase joint modeling, random interference fading and noise are adaptively suppressed, and the defading, high-fidelity, phase-continuous complex envelope signal and its corresponding continuous phase sequence are output.

[0120] Finally, based on the continuous phase sequence, phase difference demodulation is performed in the spatial dimension to extract the vibration response distributed along the optical fiber. The vibration event is accurately located by combining the optical flight time and optical path mapping relationship. At the same time, the event type is identified by time-frequency analysis or statistical feature extraction.

[0121] This method is particularly suitable for distributed acoustic and ultrasonic monitoring scenarios under long-distance, high-noise, and deep-fading conditions. It can significantly improve anti-fading capability and demodulation robustness while taking into account high spatial resolution and high system bandwidth. It has important engineering value for applications such as pipeline leak monitoring, industrial equipment structural health monitoring, and perimeter security.

[0122] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-resolution distributed acoustic-vibration demodulation method based on swept-frequency optical depth anti-fading, characterized in that, Includes the following steps: Step 1: Inject continuous linear sweep frequency probe light into the optical fiber and obtain the original back Rayleigh scattering signal containing information distributed along the optical fiber through coherent beat frequency; Step 2: Perform frequency domain demultiplexing and matched filtering pulse compression on the original backscattered Rayleigh signal to reconstruct the single-sweep signal into multiple high-resolution time-domain complex envelope signals; Step 3: Input the high-resolution time-domain complex envelope signal into the deep anti-fading demodulation network, and achieve adaptive suppression of random fading and noise of the time-domain complex envelope signal through amplitude and phase joint modeling, and output the defading phase sequence; Step 4: Demodulate the phase difference using the defading phase sequence and combine it with the optical time-of-flight positioning model to achieve demodulation and spatial positioning of vibration events.

2. The distributed acoustic-vibration demodulation method as described in claim 1, characterized in that, Step 1 specifically includes: Step 1.1: A narrow linewidth laser is used as the light source to output continuous light. The linear frequency modulated electrical signal generated by the arbitrary waveform generator is input into the electro-optic modulator to modulate the frequency of the continuous light, so that the laser frequency scans linearly with time and oscillates between the preset start frequency and end frequency, so that each instantaneous frequency point corresponds to a unique equivalent optical path. Step 1.2: The modulated sweep probe light is split by an optical fiber coupler and injected into the optical fiber under test. The probe light propagates unidirectionally along the optical fiber and undergoes Rayleigh scattering with randomly distributed scattering centers in the fiber core. Under the action of ultrasonic events or vibration disturbances, phase modulation related to local strain changes is generated. Among them, the back Rayleigh scattering component propagating in the opposite direction of incident carries the scattering intensity distribution and phase disturbance information at each position along the line, and is coupled back to the detection optical path at the optical fiber port. Step 1.3: Another part of the continuous light output from the laser is used as the local oscillator light and directly introduced into the balanced detector through an independent optical path to interfere with the back Rayleigh scattered light from the fiber under test. Since the probe light is linearly swept, a time-varying difference frequency signal is formed between the local oscillator light and the back Rayleigh scattered light. The balanced detector performs differential detection on the intensity of the two light paths and outputs an electric domain beat frequency interference signal with amplitude representing the scattering intensity and phase representing the fiber strain disturbance. Step 1.4: The electric domain beat frequency interference signal output by the balanced detector is pre-amplified and anti-aliasing filtered before being sent to the data acquisition card. Synchronous sampling is performed at a sampling rate of not less than twice the upper limit of the beat frequency. Under the multi-channel architecture, the beat frequency signals of multiple probe lights or multiple optical fibers are acquired in parallel to obtain the time-continuous original back Rayleigh scattering signal.

3. The distributed acoustic-vibration demodulation method as described in claim 1, characterized in that, Step 2 specifically includes: Step 2.1: Demultiplex the original backscattered Rayleigh signal in the frequency domain according to the frequency sweep rule: First, based on the different effective length segments in the optical fiber corresponding to different frequency intervals, the original backscattered Rayleigh signal represented in the frequency domain is divided into several sub-frequency modulated pulse frequency domain signals according to the frequency interval using this mapping. Then, a fast inverse Fourier transform is performed on each sub-frequency modulated pulse frequency domain signal to obtain the time domain signal of the sub-frequency modulated pulse. Step 2.2: Based on linear frequency modulation matched filtering technology, the preset sub-pulse signal is deconvoluted and conjugated, and then convolved with the time domain signal of each sub-frequency modulation pulse to perform pulse compression, and output a high-resolution time domain complex envelope signal sequence after compression.

4. The distributed acoustic-vibration demodulation method as described in claim 1, characterized in that, Step 3 specifically includes: Step 3.1: Set the extent of the local data block within the 3D window; Step 3.2: Divide the multi-channel high-resolution time-domain complex envelope signal into several local data blocks according to the spatial-temporal dimension, and then input them into the deep learning anti-fading demodulation network to complete feature initialization, coarse feature extraction, fine feature reconstruction and complex domain amplitude and phase recovery in sequence, and train to output the defading phase sequence.

5. The distributed acoustic-vibration demodulation method as described in claim 4, characterized in that, In step 3.2, a low-level feature description of the local data block is established through shallow 3D convolution, thus completing feature initialization.

6. The distributed acoustic-vibration demodulation method as described in claim 4, characterized in that, In step 3.2, multi-scale dilated convolution is used to capture the fading structure and weak vibration texture of the initial features, thus completing the coarse feature extraction.

7. The distributed acoustic-vibration demodulation method as described in claim 4, characterized in that, In step 3.2, the details of deep fading regions in coarse features are repaired by using a three-dimensional encoding-decoding structure in conjunction with skip connections, and the defading complex envelope is obtained through linear convolution mapping to complete the fine feature reconstruction. The three-dimensional encoding-decoding structure consists of multi-layer two-dimensional convolution operations, nonlinear activation operations, and feature normalization processing.

8. The distributed acoustic-vibration demodulation method as described in claim 7, characterized in that, In step 3.2, the defading complex envelope and the real high signal-to-noise ratio complex envelope are used together for loss construction. The amplitude reconstruction error and phase consistency error are combined to perform reverse optimization on the deep learning anti-fading demodulation network. Finally, the defading complex envelope and defading phase sequence of each local data block are seamlessly spliced ​​along the spatial and temporal dimensions to generate the final three-dimensional defading complex envelope covering the entire fiber length and the corresponding continuous phase sequence.

9. The distributed acoustic-vibration demodulation method as described in claim 1, characterized in that, Step 4 specifically includes: Step 4.1: By performing phase difference demodulation on adjacent spatial sampling points of the defading phase sequence, the common phase component is suppressed and the local disturbance response is highlighted, and the vibration response at each position along the optical fiber is extracted. Step 4.2: Perform energy statistics on the vibration response within the time sliding window, construct event judgment features, and when the energy exceeds the preset threshold, determine that there is a vibration event in the corresponding spatial location and time window, and then output the spatiotemporal location set of candidate events and their corresponding energy features. Step 4.3: Combining the optical time of flight and optical path mapping relationship, map the time index corresponding to the event to the actual optical fiber spatial location, and then determine the precise location of the event through the energy peak; Step 4.4: Perform time-frequency analysis or statistical feature extraction on the local vibration response after positioning, and output the event type and its location.

Images