A das space-time signal enhancement method for vehicle trajectory tracking

Abstract

The application provides a DAS space-time signal enhancement method for vehicle trajectory tracking, and relates to the technical field of distributed optical fiber acoustic wave sensing signal processing. The application obtains a vehicle vibration observation signal collected by a DAS system, constructs a parameterized mapping function, and takes a reconstruction fidelity term and a wavelet basis orthogonal regularization term as optimization constraints; the vehicle vibration observation signal is input into a U-shaped coding and decoding framework, multi-stage cascaded sparse decomposition is carried out through a mixed learnable wavelet transform layer, and approximate coefficients and detail coefficients under different scales are obtained; the detail coefficients are input into an enhanced residual module of a fusion channel attention mechanism, and enhanced detail features are obtained; multi-scale aggregation is carried out on deep low-frequency components and the enhanced detail features, and a denoising reconstruction signal is obtained through a parameterized reconstruction head and a non-negative projection operator. The application can improve the denoising reconstruction stability of a vehicle trajectory signal under non-stationary background noise interference.

Description

Technical Field

[0001] This invention relates to the field of distributed fiber optic acoustic wave sensing signal processing technology, and in particular to a DAS spatiotemporal signal enhancement method for vehicle trajectory tracking. Background Technology

[0002] Distributed fiber optic acoustic wave (DAS) sensing technology can detect external vibrations using backscattered Rayleigh signals along the sensing fiber, featuring long-distance continuous measurement and multi-location synchronous acquisition. He Zuyuan and Liu Qingwen, in their article "Principles and Applications of Fiber Optic Distributed Acoustic Wave Sensors" published in *Laser & Optoelectronics Progress*, Vol. 58, No. 13, 2021, describe how DAS sensors can simultaneously acquire waveforms of mechanical vibrations at multiple locations along a fiber optic link with a high sampling rate. When a vehicle travels on a road, the vibrations generated by the tires contacting the road surface propagate along the roadbed and surrounding medium, acting on the sensing fiber near the road. Wang Maoning et al., in their article "Vehicle Trajectory Extraction Method Based on Distributed Fiber Optic Sensing System" published in *Engineering Science & Technology*, Vol. 53, No. 2, 2021, describe how a distributed fiber optic sensing system can acquire the spatiotemporal response formed when a vehicle passes by, and further extract the vehicle trajectory. These studies demonstrate that DAS-acquired data can already serve as a data source for vehicle trajectory tracking.

[0003] Existing technologies for DAS vehicle signal processing typically require preprocessing of the acquired signals before extracting features related to vehicle events from the time domain, frequency domain, or time-frequency domain. Chinese patent document CN107256635A discloses a fiber optic sensing traffic monitoring method based on phase-sensitive optical time-domain reflectometry. This method acquires vehicle vibration signals through a distributed fiber optic sensing system and identifies vehicle location, quantity, and type using a combination of time-domain, frequency-domain, and time-frequency analysis. Addressing the issues of background disturbances and weak signals in DAS data, the 2025 issue of *Progress in Geophysics*, titled "Research Progress on Denoising Methods and Applications of Distributed Acoustic Sensing Data," reviews DAS data denoising methods, covering approaches such as frequency domain filtering, wavelet transform, and deep learning denoising. It is evident that DAS vehicle trajectory tracking does not solely depend on signal acquisition accuracy; post-acquisition noise suppression, feature preservation, and signal reconstruction also directly affect the trajectory extraction results.

[0004] In the field of intelligent processing of distributed optical fiber sensor signals, Chinese patent document CN113049084B discloses a distributed optical fiber sensor signal recognition method based on an attention mechanism and a ResNet. This method first preprocesses typical event signals and constructs a time-frequency feature dataset, then uses a residual network with an attention mechanism to identify the event to be measured. This type of method demonstrates that attention mechanisms and residual networks can already be used for feature representation of distributed optical fiber sensor signals. However, DAS signal enhancement required for vehicle trajectory tracking focuses more on the reconstruction quality of noisy observation signals to vehicle vibration signals. Existing processing methods still need to further improve the denoising reconstruction effect of vehicle trajectory signals under non-stationary noise interference. Summary of the Invention

[0005] The purpose of this invention is to provide a DAS spatiotemporal signal enhancement method for vehicle trajectory tracking, which can solve the problem of reduced reconstruction quality of vehicle vibration signal under non-stationary background noise interference, and achieve multi-scale denoising reconstruction and stable enhancement effect of vehicle trajectory signal under energy non-negativity constraint.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A DAS spatiotemporal signal enhancement method for vehicle trajectory tracking includes:

[0008] The vehicle vibration observation signal collected by the DAS system is acquired. The vehicle vibration observation signal is formed by superimposing the true value of the vehicle vibration signal and the background noise signal. The true value of the vehicle vibration signal satisfies the non-negativity constraint.

[0009] A parameterized mapping function is constructed based on vehicle vibration observation signals, and the reconstruction fidelity term and wavelet basis orthogonal regularization term are used as optimization constraints to obtain a reconstruction mapping for approximating the true value of vehicle vibration signals from vehicle vibration observation signals.

[0010] Vehicle vibration observation signals are input into a U-shaped encoder-decoder architecture based on wavelet decomposition multi-resolution analysis. In the encoding stage, the vehicle vibration observation signals are subjected to multi-level cascaded sparse decomposition through a hybrid learnable wavelet transform layer. The hybrid learnable wavelet transform layer includes a fixed wavelet basis and an adaptive wavelet decomposition module. The fixed wavelet basis is used to capture the basic transient structure, and the adaptive wavelet decomposition module is used to generate low-pass filter weights and high-pass filter weights according to the local statistical characteristics of the input signal, so as to obtain approximation coefficients and detail coefficients at different scales.

[0011] The detail coefficients are input into the enhanced residual module with the fusion channel attention mechanism. The wavelet domain features are enhanced and the weights are recalibrated through depthwise separable convolution and SE channel attention mechanism to obtain enhanced detail features.

[0012] In the decoding stage, the deep low-frequency components and the corresponding scale-enhanced detail features are aggregated at multiple scales to obtain a global aggregated representation. The global aggregated representation is then input into the parameterized reconstruction head for reconstruction mapping. A non-negative projection operator is introduced at the end of the mapping of the parameterized reconstruction head to obtain a denoised reconstructed signal that conforms to the non-negative physical properties of DAS vibration signal energy.

[0013] Preferably, the vehicle vibration observation signal acquired by the DAS system is obtained. The vehicle vibration observation signal is formed by superimposing the true value of the vehicle vibration signal and the background noise signal. The true value of the vehicle vibration signal satisfies a non-negativity constraint, including:

[0014] The vibration response generated by the vehicle acting on the sensing fiber during its road travel is taken as the true value of the vehicle vibration signal.

[0015] The DAS background interference affected by the scale parameter is used as the background noise signal;

[0016] The vehicle vibration signal is superimposed with the background noise signal to form the vehicle vibration observation signal;

[0017] The true value of the vehicle vibration signal is defined as a non-negative signal so that the true value of the vehicle vibration signal satisfies the physical characteristic of non-negative energy of the DAS vibration signal.

[0018] Preferably, a parameterized mapping function is constructed based on the vehicle vibration observation signal, and the reconstruction fidelity term and the wavelet basis orthogonal regularization term are used as optimization constraints to obtain a reconstruction mapping for approximating the true value of the vehicle vibration signal from the vehicle vibration observation signal, including:

[0019] Establish a parameterized mapping function that takes vehicle vibration observation signals as input and reconstructed signals as mapping results;

[0020] The time-domain Euclidean distance between the reconstructed signal and the target waveform is constrained by a reconstruction fidelity term.

[0021] The filter bank matrix consisting of low-pass and high-pass filters constrained by wavelet basis orthogonal regularization terms satisfies the orthogonal basis condition;

[0022] The parameterized mapping function is optimized based on the reconstruction fidelity term and the wavelet basis orthogonal regularization term to form the reconstruction mapping.

[0023] Preferably, during the encoding stage, the vehicle vibration observation signal is subjected to multi-level cascaded sparse decomposition through a hybrid learnable wavelet transform layer, including:

[0024] Use vehicle vibration observation signals as the first-layer input signals;

[0025] By performing one-dimensional convolution and downsampling on the input signal using a low-pass filter in the learnable wavelet transform layer, the approximation coefficients for the next level are obtained.

[0026] By performing one-dimensional convolution and downsampling on the input signal through a high-pass filter in the learnable wavelet transform layer, the detail coefficients of the next layer are obtained.

[0027] The approximation coefficients of the next level are used as the input signals of subsequent levels for cascaded decomposition to obtain approximation coefficients and detail coefficients at different scales.

[0028] Preferably, the hybrid learnable wavelet transform layer includes a fixed wavelet basis and an adaptive wavelet decomposition module. The fixed wavelet basis is used to capture the basic transient structure, and the adaptive wavelet decomposition module is used to generate low-pass filter weights and high-pass filter weights based on the local statistical characteristics of the input signal, including:

[0029] In the shallow decomposition stage, the input signal is decomposed by fixing the wavelet basis in order to capture the basic transient structure in the input signal;

[0030] In the deep decomposition stage, the local statistical characteristics of the input signal are extracted through the adaptive wavelet decomposition module;

[0031] Low-pass filter weights and high-pass filter weights that match the input signal are generated based on local statistical characteristics;

[0032] Sparse representation of non-stationary time-varying features in deep feature space is achieved by using low-pass filter weights and high-pass filter weights.

[0033] Preferably, the local statistical properties of the input signal are extracted using an adaptive wavelet decomposition module, including:

[0034] Global average pooling is performed on the input signal to obtain the global context vector;

[0035] The global context vector is input into the fully connected layer for nonlinear mapping to obtain the dynamic filter generation parameters;

[0036] Generate low-pass filter weights and high-pass filter weights based on dynamic filter generation parameters;

[0037] Apply a non-negative truncation constraint to the input signal, and then perform convolution operations between the input signal with the weights of the low-pass filter and the weights of the high-pass filter, respectively.

[0038] The absolute value activation is applied to the convolution result to obtain the approximation coefficients and detail coefficients.

[0039] Preferably, the detail coefficients are input into the enhanced residual module with a fusion channel attention mechanism. The wavelet domain features are then enhanced and weights are recalibrated using depthwise separable convolution and SE channel attention mechanisms to obtain enhanced detail features, including:

[0040] Use detail coefficients as input features for the enhanced residual module;

[0041] Spatial features are extracted from the input features using depthwise separable convolution to obtain convolutional features;

[0042] Channel dependencies of convolutional features are modeled using the SE channel attention mechanism to obtain channel weights;

[0043] The convolutional features are recalibrated based on the channel weights to obtain recalibrated features.

[0044] The input features and recalibration features are residually fused, and the residual fusion result is nonlinearly activated to obtain enhanced detail features.

[0045] Preferably, during the decoding stage, multi-scale aggregation is performed on the deep low-frequency components and the corresponding scale-enhanced detail features to obtain a global aggregated representation, including:

[0046] Deep low-frequency components are used as the initial reconstruction features in the decoding stage;

[0047] The initial reconstructed features are upsampled to obtain the upsampled reconstructed features;

[0048] The upsampled reconstructed features and the enhanced detail features of the corresponding scale are concatenated along the channel dimension to obtain the aggregated features at the current scale;

[0049] The aggregated features at the current scale are passed down to the shallow layers as the reconstructed features at the next scale until a global aggregated representation is obtained.

[0050] Preferably, the global aggregation representation is input to the parameterized reconstruction head for reconstruction mapping, and a non-negative projection operator is introduced at the end of the mapping of the parameterized reconstruction head to obtain a denoised reconstruction signal that conforms to the non-negative physical characteristics of DAS vibration signal energy, including:

[0051] Input the global aggregate representation into the parameterized reconstruction head;

[0052] The initial reconstructed signal is obtained by performing channel compression and temporal waveform reconstruction on the global aggregate representation through multi-layer convolutional mapping in the parameterized reconstruction head;

[0053] At the end of the mapping of the parameterized reconstruction head, the negative components in the initial reconstructed signal are filtered out by the non-negative projection operator;

[0054] The initial reconstructed signal after filtering out negative components is used as the denoised reconstructed signal.

[0055] Preferably, the parameterized mapping function is optimized based on the reconstruction fidelity term and the wavelet basis orthogonal regularization term to form a reconstruction mapping, including:

[0056] Construct a hybrid dataset consisting of relatively clean real signals and simulated synthetic data;

[0057] Generating true samples of vehicle vibration signals based on a mass spring damping dynamic model of the vehicle-road coupled system;

[0058] The true value sample of vehicle vibration signal is decomposed into transient impact component and quasi-periodic oscillation component;

[0059] Construct a heterogeneous noise environment that includes colored noise and Rayleigh distributed multiplicative noise;

[0060] The true value samples of vehicle vibration signals and heterogeneous noise environment are superimposed to form the vibration observation signal of the training vehicle.

[0061] The parameterized mapping function is optimized using vibration observation signals from training vehicles, so that the optimized parameterized mapping function forms a reconstructed mapping.

[0062] The present invention discloses the following beneficial effects:

[0063] This invention limits the vehicle vibration observation signal acquired by the DAS system to the superposition of the true value of the vehicle vibration signal and the background noise signal, and ensures that the true value of the vehicle vibration signal satisfies the non-negativity constraint. This allows subsequent reconstruction processing to be based on the physical properties of the DAS vibration signal, avoiding signal components in the reconstruction result that do not conform to the non-negativity of energy. By constructing a parameterized mapping function and using reconstruction fidelity and wavelet basis orthogonality regularization terms as optimization constraints, the vehicle vibration observation signal is simultaneously constrained by waveform similarity and wavelet basis orthogonality when approximating the true value of the vehicle vibration signal, reducing reconstruction deviation caused by direct fitting of noisy observation signals. By setting a hybrid learnable wavelet transform layer in the encoding stage of the U-shaped encoder-decoder architecture, the basic transient structure is captured using a fixed wavelet basis, and an adaptive wavelet decomposition module generates low-pass filter weights and high-pass filter weights based on the local statistical characteristics of the input signal. The weighting of the wavelet coefficients decomposes the vehicle vibration observation signal into approximation coefficients and detail coefficients at different scales, addressing the problem of insufficient adaptability of fixed wavelet bases to non-stationary noise. By inputting the detail coefficients into the enhanced residual module with a fusion channel attention mechanism, wavelet domain feature enhancement and weight recalibration are completed using depthwise separable convolution and SE channel attention mechanism. This preserves the features in the detail coefficients that are strongly correlated with the vehicle's transient vibration, while reducing the redundant channel weights corresponding to background noise. By performing multi-scale aggregation of deep low-frequency components and corresponding enhanced detail features during the decoding stage, and introducing a non-negative projection operator at the end of the mapping of the parameterized reconstruction head, the reconstructed denoised signal contains both low-frequency main information and multi-scale detail information, while satisfying the physical property of non-negative energy of DAS vibration signals. This improves the stability of denoising and reconstructing vehicle trajectory signals under non-stationary background noise interference. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0066] Figure 2 This is a schematic diagram of the backbone architecture of the denoising model provided in an embodiment of the present invention;

[0067] Figure 3 This is a schematic diagram of the adaptive wavelet decomposition module structure provided in an embodiment of the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this embodiment provides a DAS spatiotemporal signal enhancement method for vehicle trajectory tracking, including:

[0071] Step 100: Acquire the vehicle vibration observation signal collected by the DAS system; the vehicle vibration observation signal is formed by superimposing the true value of the vehicle vibration signal and the background noise signal, and the true value of the vehicle vibration signal satisfies the non-negativity constraint;

[0072] Step 200: Construct a parameterized mapping function based on the vehicle vibration observation signal, and use the reconstruction fidelity term and the wavelet basis orthogonal regularization term as optimization constraints to obtain the reconstruction mapping used to approximate the true value of the vehicle vibration signal from the vehicle vibration observation signal;

[0073] Step 300: Input the vehicle vibration observation signal into a U-shaped encoder-decoder architecture based on wavelet decomposition multi-resolution analysis. In the encoding stage, the vehicle vibration observation signal is subjected to multi-level cascaded sparse decomposition through a hybrid learnable wavelet transform layer. The hybrid learnable wavelet transform layer includes a fixed wavelet basis and an adaptive wavelet decomposition module. The fixed wavelet basis is used to capture the basic transient structure, and the adaptive wavelet decomposition module is used to generate low-pass filter weights and high-pass filter weights according to the local statistical characteristics of the input signal, so as to obtain approximation coefficients and detail coefficients at different scales.

[0074] Step 400: Input the detail coefficients into the enhanced residual module with the fusion channel attention mechanism, and perform feature enhancement and weight recalibration on the wavelet domain features through depthwise separable convolution and SE channel attention mechanism to obtain enhanced detail features;

[0075] Step 500: In the decoding stage, multi-scale aggregation is performed on the deep low-frequency components and the enhanced detail features of the corresponding scale to obtain a global aggregation representation. The global aggregation representation is input into the parameterized reconstruction head for reconstruction mapping. A non-negative projection operator is introduced at the end of the mapping of the parameterized reconstruction head to obtain a denoised reconstruction signal that conforms to the non-negative physical characteristics of DAS vibration signal energy.

[0076] In step 100, the vehicle vibration observation signal acquired by the DAS system is a noisy observation signal. Considering the physical characteristics of the DAS system and the influence of coherent fading, the vehicle vibration observation signal is defined as an additive noise model with non-negative constraints:

[0077] (1)

[0078] In equation (1), This represents the true value of the vehicle vibration signal. Background noise signal affected by scale parameters, The vehicle vibration observation signals acquired by the DAS system The length of the true value of the vehicle vibration signal corresponds to the vibration response generated by the vehicle acting on the sensing fiber during road travel. The background noise signal corresponds to the background interference superimposed on the vehicle vibration response during the acquisition process of the DAS system. The non-negativity constraint is used to limit the range of values ​​of the true value of the vehicle vibration signal, so that the signal obtained by subsequent reconstruction conforms to the physical characteristic of non-negativity of DAS vibration signal energy.

[0079] The reconstruction of the vehicle vibration observation signal is an ill-posed problem. When directly solving for the true value of the vehicle vibration signal from the observation signal, background noise can affect the reconstruction stability and may weaken the transient components of the vehicle trajectory signal. To address this issue, a parameterized mapping function is constructed in step 200. , to reconstruct the signal Approximating the real signal while satisfying physical constraints To find the optimal approximation in the solution space, this problem is transformed into a minimization problem of a regularized functional:

[0080] (2)

[0081] In equation (2), To reconstruct the fidelity term, For the orthogonal regularization term of the wavelet basis, For balance coefficient, These are the learnable parameters of the parameterized mapping function. The reconstruction fidelity term is used to constrain the temporal differences between the reconstructed signal and the target waveform. The orthogonal regularization term of the wavelet basis is used to constrain the orthogonality of the filter bank used in wavelet decomposition and reduce energy distribution anomalies caused by filter coefficient degradation.

[0082] In step 300, the vehicle vibration observation signal input is based on a U-shaped encoder-decoder architecture using wavelet decomposition multi-resolution analysis. This U-shaped encoder-decoder architecture employs a hybrid learnable wavelet transform layer as the downsampling operator, replacing the pooling operation in traditional convolutional neural networks, which is prone to causing loss of high-frequency details. Low-pass and high-pass filter coefficients are dynamically generated through parameterized subnetworks based on the local statistical characteristics of the input signal, achieving data-driven time-frequency sparse representation.

[0083] like Figure 2 As shown, the vehicle vibration observation signal is used as the Input Signal, and the input dimension can be represented as follows: , This represents the batch dimension. The encoding stage decomposes progressively downwards along the time dimension, passing through DaubechiesWavelet, Morlet Wavelet, and multiple Adaptive Wavelets to obtain multi-scale Low Frequency and High Frequency. High Frequency features are enhanced and projected through their corresponding ResBlock+ProjOut layers and then passed to the decoding stage via a Skip path. The decoding stage uses Concat Level 5, Concat Level 4, Concat Level 3, Concat Level 2, and Concat Level 1. Each Concat level is used to concatenate the upsampled reconstructed features with the corresponding high-frequency enhanced features. The shallowest aggregation layer yields a dimension of... The features are then input into the Reconstruction Head. The Reconstruction Head includes multiple layers of convolutional mappings. Figure 2 The channel dimension is shown in the figure. Mapped to Then by Mapped to Finally by Mapped to The dimension is obtained as The denoised output.

[0084] Figure 2 The right side shows the internal structure of the enhanced residual module. The enhanced residual module is structured with dimension... The features are used as input, and sequentially processed through Depthwise Conv, Pointwise Conv, SE-Attention, and Element-wise Add to obtain a result with the same dimensionality. The output consists of: Depthwise Conv for extracting local spatial features within each channel, Pointwise Conv for inter-channel mapping, SE-Attention for establishing channel dependencies and generating channel weights, and Element-wise Add for performing a Residual Connection between the input features and the attention-enhanced features. Residual connections are used to enhance wavelet domain detail features while preserving effective transient information in the original detail components.

[0085] By learning parameterized filter banks Achieve multi-resolution wavelet decomposition. For the first Hierarchical input signals Decompose the material to obtain the approximate coefficients for the next level. and detail coefficient :

[0086] (3)

[0087] In equation (3), This represents a one-dimensional convolution operator. Indicates the downsampling operator. Indicates the first Hierarchical low-pass filter, Indicates the first Hierarchical high-pass filters, Indicates the first Hierarchical input signals, This represents the approximation coefficient for the next level. This represents the detail coefficient of the next level.

[0088] The filter bank of equation (3) Defined as:

[0089] (4)

[0090] In equation (4), Represents a fixed set of wavelet bases. For adaptive generative networks, The parameters are for the adaptive generative network. The first two decomposition levels use a fixed wavelet basis to capture the basic transient structure. The third and subsequent decomposition levels use an adaptive generative network. Low-pass and high-pass filters are generated based on the input signal to adapt to non-stationary time-varying features in the deep feature space.

[0091] In one embodiment, the encoder module employs a five-stage cascaded sparse decomposition strategy to progressively map the high-dimensional time-domain signal to a compact latent feature space. The first two stages of decomposition introduce a fixed wavelet basis to capture the basic transient structure. The first stage utilizes the Daubechies 4 wavelet to downsample the time dimension to... The output high-frequency feature dimension is The second stage uses Morlet wavelets to further compress the time dimension to... The generation dimension is The feature tensor. To overcome the problem that fixed basis function prior constraints are insufficient for representing non-stationary time-varying features, the network embeds adaptive wavelet decomposition modules starting from the third level. As the layers deepen, the adaptive modules gradually encode deep semantic information, and the temporal resolution decreases sequentially. , and The corresponding feature channel numbers are shrunk to , and The encoder obtains a dimension of [dimensional value] at the bottleneck layer. The deep low-frequency components.

[0092] like Figure 3As shown, the adaptive wavelet decomposition module includes Input, Global Pool, FC Layers, Filter hHigh, Filter gLow, Clamp, Dyn.Conv, Abs, Output Low Freq, and Output High Freq. The input features are firstly processed by Global Pool for global average pooling to obtain a global context vector. This global context vector is then processed by FC Layers for nonlinear mapping, generating dynamic filter weights for Filter hHigh and Filter gL, respectively. Secondly, the input features are processed by Clamp for non-negative truncation constraints, resulting in input features that satisfy these constraints. The dynamic filter weights serve as the weights input to Dyn.Conv, and the input features processed by Clamp serve as the data input to Dyn.Conv, performing dynamic convolution operations on the high-frequency and low-frequency branches, respectively. The dynamic convolution results of the high-frequency and low-frequency branches are then activated by Abs, yielding Output High Freq and Output Low Freq, respectively. Output High Freq corresponds to detail coefficients, and Output Low Freq corresponds to approximation coefficients. Figure 3 The annotation dimensions for Input, Global Pool, FCLayers, Clamp, Dyn.Conv, and Abs are used to illustrate the batch dimension of the adaptive wavelet decomposition module. Channel Dimension and sequence length Data transfer relationships.

[0093] In step 400, the detail coefficients are input into the enhanced residual module of the fusion channel attention mechanism. This is for the detail components in the wavelet domain. To address problems involving noise and redundant information, an enhanced residual operator is introduced. Perform high-dimensional manifold projection and feature recalibration:

[0094] (5)

[0095] In equation (5), Indicates the first The level of detail, Indicates the first Enhanced detail features at different levels, This indicates an enhanced residual operator. The parameters represent the enhanced residual operator. Represents depthwise separable convolution. For channel attention operators, This represents a non-linear activation function.

[0096] The enhanced residual module utilizes depthwise separable convolutions and an inverted bottleneck structure to capture spatial features while maintaining channel independence. The enhanced residual module utilizes a dilation rate of... Pointwise convolution maps features to a high-dimensional manifold space, improving the separability of wavelet domain features. The SE channel attention mechanism achieves adaptive recalibration of feature weights by explicitly modeling inter-channel dependencies. After recalibration, redundant noise channels dominated by coherent fading are suppressed, while signal channels corresponding to key transient information are enhanced, resulting in improved signal-to-noise ratio of the feature map.

[0097] In step 500, the decoder employs a progressive aggregation strategy, fusing deep abstract semantics and shallow physical details across scales to recover the full-resolution temporal waveform of the signal. To compensate for the spatial resolution loss caused by downsampling and to fuse multi-scale information, a cascaded upsampling aggregation path is designed. The aggregated feature map is shown below. Calculated using the following recursive equation:

[0098] (6)

[0099] In equation (6), This represents the bilinear interpolation upsampling operator. This represents the feature concatenation operator along the channel dimension. Indicates the first Hierarchical reconstruction features Indicates the first Hierarchical aggregation features Indicates the first Layered enhancement of detail features. The decoding stage begins reconstruction from the deepest layer. Deep low-frequency components are first concatenated with enhancement of detail features at the same layer, and then upsampled. The upsampled features continue to be concatenated with enhancement of detail features at the previous layer until a global aggregate representation is formed.

[0100] In one embodiment, the decoding phase starts from dimension 1. The deep low-frequency components begin to be reconstructed. These deep low-frequency components are then stitched together with detail features of the same level to generate... Fuse tensors and upsample to The upsampled features are then concatenated with the fourth layer of detail features, restoring the output dimension to its original value. This process is executed recursively upwards along the hierarchy, with the number of feature tensor channels and sequence length recovered layer by layer. The aggregated output of the third layer is... The second-level aggregation output is Finally, it is concatenated with the high-frequency features of the first layer to form a dimension of The global aggregate representation is input to the parameterized reconstruction head, and the channel dimension is compressed through multiple convolutional mappings to obtain a dimension of [dimensionality missing]. The denoised and reconstructed signal.

[0101] To ensure the algorithm network maintains high fidelity under non-stationary noise interference, a multi-task collaborative optimization loss function is defined to quantitatively evaluate the parameterized reconstruction head. Overall loss function. As a mathematical standard for evaluating the quality of reconstruction, it is defined as follows:

[0102] (7)

[0103] In equation (7), This indicates the reconstruction fidelity term. This represents the wavelet orthogonal regularization term. These are adaptive weighting coefficients used to balance reconstruction accuracy and basis function completeness. Guide the network to learn nonlinear mappings to recover transient features and ensure the reconstruction of waveforms. With target waveform The Euclidean distance is minimized in the time domain, as defined below:

[0104] (8)

[0105] In equation (8), Indicates the number of samples. Indicates the first The reconstructed waveform of each sample, Indicates the first The target waveform of each sample This represents the L2 norm.

[0106] Meanwhile, to avoid trivial solutions in the learned filter coefficients and to ensure the conservation of signal energy during multi-resolution decomposition, a wavelet orthogonal regularization term is introduced. This term forces the generation of the low-pass filter. With high-pass filter The resulting filter matrix satisfies the orthogonal basis condition, defined as follows:

[0107] (9)

[0108] In equation (9), Represents the filter bank matrix. It is the identity matrix. This represents the Frobenius norm.

[0109] Under the supervision of the aforementioned loss function, the final signal is reconstructed through a parameterized reconstruction head. Complete. To ensure the output signal conforms to the physical characteristic of non-negative energy of DAS vibration signals, a hard threshold constraint is introduced at the end of the mapping. The final reconstructed signal is defined as follows:

[0110] (10)

[0111] In equation (10), It is a non-linear mapping function. It is a non-negative projection operator. This represents the global aggregate representation of the input parameterized reconstruction head. Indicates the first Each convolutional kernel weight, Indicates the first One bias term, Represents the ReLU activation function. "in "" indicates the convolutional mapping input that enters the ReLU activation function in the parameterized reconstruction header. This represents the channel or feature component index in the convolutional mapping. Using the non-negative projection operator, the trained reconstruction mapping head filters out negative components that do not conform to physical common sense during output, achieving consistent convergence from mathematical optimization to physical logic.

[0112] In one embodiment, to address the issue that real DAS observations cannot obtain pure, noise-free vehicle spatiotemporal signals, a hybrid dataset driven by physical mechanisms is constructed. Raw DAS data contains system circuit noise, and there is a significant imbalance between the acquired signal and the clean signal. Traditional research using sinusoidal simulations struggles to cover vehicle-road coupling conditions, limiting the generalization ability of data-driven models in practical applications. The hybrid dataset consists of selected relatively clean real signals and simulated synthetic data. It comprises measured samples from the Chang'an University Vehicle Networking and Intelligent Vehicle Test Site and synthetic data based on dynamic simulation, covering both real dynamic characteristics and the non-stationary noise distribution unique to DAS. The hybrid dataset is divided into a training set of 800 samples and a test set of 200 samples.

[0113] DAS observation signals of synthetic data It can be regarded as a vehicle spatiotemporal signal With heterogeneous environmental noise Nonlinear superposition:

[0114] (11)

[0115] Considering the complex vehicle-road coupling mechanism, the "vehicle-road" coupled system is abstracted into a classic mass-spring-damped dynamics model. According to Newton's second law, the vertical displacement response excited by the vehicle... It obeys the following second-order linear differential equation:

[0116] (12)

[0117] In equation (12), , , These represent the equivalent mass, damping coefficient, and stiffness, respectively. This represents the external action triggered by the vehicle. Based on the general solution of the dynamic equation, the vehicle signal is decoupled into a transient impact component and a quasi-periodic oscillation component generated by the tire-road interaction. The transient impact component is fitted using a Ricker wavelet, and the quasi-periodic oscillation component is fitted using a damped sine function to capture the time-domain impact attenuation and frequency-domain resonance characteristics of the real signal.

[0118] To address the unique non-stationary interference problem of DAS systems, a system is constructed. Heterogeneous noise environment consisting of colored noise and Rayleigh distributed multiplicative noise. Colored noise is used to fit environmental thermal drift and geological microseismic activity. Rayleigh multiplicative noise is used to fit fiber coherent fading effects. By superimposing vehicle signals with heterogeneous noise environments to generate synthetic DAS observation signals, the lack of real training data is compensated for, enhancing the generalization ability and robustness of the denoising model under complex real-world conditions.

[0119] Through the above implementation methods, this embodiment decomposes the vehicle vibration observation signal into approximation coefficients and detail coefficients at different scales based on the hybrid cascaded decomposition of fixed wavelet and adaptive wavelet; based on the enhanced residual module of the fusion channel attention mechanism, the wavelet domain detail coefficients are enhanced in features and recalibrated in weights; based on the cascaded upsampling aggregation path, the deep low-frequency components and multi-scale enhanced detail features are spliced ​​and aggregated; based on the parameterized reconstruction head and non-negative projection operator, the global aggregated representation is mapped into a denoised reconstructed signal that conforms to the non-negative physical properties of DAS vibration signal energy.

[0120] In another implementation, during the multi-scale aggregation of deep low-frequency components and corresponding scale-enhanced detail features in the decoding stage, a scale-energy consistency gating is further introduced. The scale-energy consistency gating is used to compare the energy matching degree between the upsampled reconstructed features and the corresponding scale-enhanced detail features. When the scale energy of the enhanced detail features is close to that of the upsampled reconstructed features, the enhanced detail features participate in channel-dimensional stitching with a higher weight; when the scale energy of the enhanced detail features deviates abnormally from that of the upsampled reconstructed features, the weight of the enhanced detail features is reduced before stitching to minimize the impact of local scale noise spikes on the temporal waveform reconstruction.

[0121] Specifically, in the first Hierarchy will reconstruct features After bilinear interpolation upsampling operator After processing, the scale energy of the upsampled reconstructed features is calculated. :

[0122]

[0123] In equation (13), Indicates the first Scale energy of features reconstructed by hierarchical upsampling. Indicates the first The number of sampling points in hierarchical features Indicates the first feature in the upsampled reconstruction. 1 eigenvalue, Indicates the first Hierarchical reconstruction features This represents the bilinear interpolation upsampling operator.

[0124] Furthermore, the first Hierarchical Enhanced Details After bilinear interpolation upsampling operator After processing, the scale energy of the enhanced detail features is calculated. :

[0125]

[0126] In equation (14), Indicates the first Scale energy for enhancing details at different levels. Indicates the first enhanced detail feature after upsampling 1 eigenvalue, Indicates the first Enhanced detail features at different levels. Equations (13) and (14) use the same number of sampling points. This ensures that energies of two different scales can be compared at the same scale.

[0127] Based on the scale energy of the upsampled reconstructed features and scale energy to enhance detailed features Determine the first Hierarchical scale-energy consistency gating coefficient :

[0128]

[0129] In equation (15), Indicates the first Hierarchical scale energy consistency gating coefficient This indicates a positive number used to avoid a denominator of zero. Because... and Both are in the form of the sum of squares and the mean, and their values ​​are non-negative. Adding... Then, the logarithmic term in equation (15) has a definite computational object. When and When approaching, The absolute value is close to , near ;when Significantly greater than or less than hour, It decreases as the degree of energy deviation increases.

[0130] Before performing channel-dimensional stitching, the scale-energy consistency gating coefficient is used. The enhanced detail features after upsampling are modulated to form the first... Hierarchical aggregation features :

[0131]

[0132] In equation (16), Indicates the first Hierarchical aggregation features This represents the feature concatenation operator along the channel dimension. This represents the enhanced detail features modulated by the scale-energy consistency gating coefficient. Equation (16) adds amplitude modulation only to the enhanced detail features based on the original multi-scale aggregation relationship, without changing the basic structure of the reconstructed features and the enhanced detail features spliced ​​along the channel dimension.

[0133] Through the above processing, the decoding stage, while preserving deep low-frequency components and corresponding scale-enhanced detail features, adaptively suppresses enhanced detail features with abnormal energy deviations. The global aggregation representation entering the parametric reconstruction head thus reduces local scale noise spikes, and the denoised reconstructed signal maintains a more stable vehicle vibration response in the time domain waveform. This implementation does not change the hybrid learnable wavelet decomposition process, the feature enhancement process of the enhanced residual module, or the non-negative projection constraint at the end of the parametric reconstruction head in the encoding stage.

[0134] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0135] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A DAS spatiotemporal signal enhancement method for vehicle trajectory tracking, characterized in that, include: The vehicle vibration observation signal collected by the DAS system is acquired. The vehicle vibration observation signal is formed by superimposing the true value of the vehicle vibration signal and the background noise signal. The true value of the vehicle vibration signal satisfies the non-negativity constraint. A parameterized mapping function is constructed based on the vehicle vibration observation signal, and the reconstruction fidelity term and the wavelet basis orthogonal regularization term are used as optimization constraints to obtain a reconstruction mapping for approximating the true value of the vehicle vibration signal from the vehicle vibration observation signal. The vehicle vibration observation signal is input into a U-shaped encoder-decoder architecture based on wavelet decomposition multi-resolution analysis. In the encoding stage, the vehicle vibration observation signal is subjected to multi-level cascaded sparse decomposition through a hybrid learnable wavelet transform layer. The hybrid learnable wavelet transform layer includes a fixed wavelet basis and an adaptive wavelet decomposition module. The fixed wavelet basis is used to capture the basic transient structure, and the adaptive wavelet decomposition module is used to generate low-pass filter weights and high-pass filter weights according to the local statistical characteristics of the input signal, so as to obtain approximation coefficients and detail coefficients at different scales. The detail coefficients are input into the enhanced residual module with the fusion channel attention mechanism. The wavelet domain features are enhanced and the weights are recalibrated through depthwise separable convolution and SE channel attention mechanism to obtain enhanced detail features. In the decoding stage, the deep low-frequency components and the corresponding enhanced detail features are aggregated at multiple scales to obtain a global aggregated representation. The global aggregated representation is then input into the parameterized reconstruction head for reconstruction mapping. A non-negative projection operator is introduced at the end of the mapping of the parameterized reconstruction head to obtain a denoised reconstructed signal that conforms to the non-negative physical characteristics of DAS vibration signal energy.

2. The DAS spatiotemporal signal enhancement method for vehicle trajectory tracking according to claim 1, characterized in that, The vehicle vibration observation signal acquired by the DAS system is obtained. This vehicle vibration observation signal is formed by superimposing the true value of the vehicle vibration signal and the background noise signal. The true value of the vehicle vibration signal satisfies a non-negativity constraint, including: The vibration response generated by the vehicle acting on the sensing fiber during its road travel is taken as the true value of the vehicle vibration signal. The DAS background interference affected by the scale parameter is used as the background noise signal; The vehicle vibration observation signal is formed by superimposing the true value of the vehicle vibration signal and the background noise signal. The true value of the vehicle vibration signal is defined as a non-negative signal so that the true value of the vehicle vibration signal satisfies the physical characteristic of non-negative energy of DAS vibration signal.

3. The DAS spatiotemporal signal enhancement method for vehicle trajectory tracking according to claim 1, characterized in that, A parameterized mapping function is constructed based on the vehicle vibration observation signal, and the reconstruction fidelity term and wavelet basis orthogonal regularization term are used as optimization constraints to obtain a reconstruction mapping for approximating the true value of the vehicle vibration signal from the vehicle vibration observation signal, including: Establish a parameterized mapping function that takes the vehicle vibration observation signal as input and the reconstructed signal as the mapping result; The reconstruction fidelity term is used to constrain the time-domain Euclidean distance between the reconstructed signal and the target waveform; The filter bank matrix composed of the low-pass filter and the high-pass filter constrained by the wavelet basis orthogonal regularization term satisfies the orthogonal basis condition; The parameterized mapping function is optimized based on the reconstruction fidelity term and the wavelet basis orthogonal regularization term to form the reconstruction mapping.

4. The DAS spatiotemporal signal enhancement method for vehicle trajectory tracking according to claim 1, characterized in that, During the encoding stage, the vehicle vibration observation signal is subjected to multi-level cascaded sparse decomposition through a hybrid learnable wavelet transform layer, including: The vehicle vibration observation signal is used as the first-layer input signal; The input signal is subjected to one-dimensional convolution and downsampling by the low-pass filter in the hybrid learnable wavelet transform layer to obtain the approximation coefficients of the next level. The input signal is subjected to one-dimensional convolution and downsampling through the high-pass filter in the hybrid learnable wavelet transform layer to obtain the detail coefficients of the next layer. The approximation coefficients of the next level are used as input signals for subsequent levels to perform cascade decomposition, resulting in approximation coefficients and detail coefficients at different scales.

5. The DAS spatiotemporal signal enhancement method for vehicle trajectory tracking according to claim 1, characterized in that, The hybrid learnable wavelet transform layer includes a fixed wavelet basis and an adaptive wavelet decomposition module. The fixed wavelet basis is used to capture the basic transient structure, and the adaptive wavelet decomposition module is used to generate low-pass filter weights and high-pass filter weights based on the local statistical characteristics of the input signal, including: In the shallow decomposition stage, the input signal is decomposed using the fixed wavelet basis to capture the fundamental transient structure in the input signal; In the deep decomposition stage, the local statistical characteristics of the input signal are extracted through the adaptive wavelet decomposition module; Based on the local statistical characteristics, low-pass filter weights and high-pass filter weights that match the input signal are generated; The low-pass filter weights and high-pass filter weights are used to perform sparse representation of non-stationary time-varying features in the deep feature space.

6. The DAS spatiotemporal signal enhancement method for vehicle trajectory tracking according to claim 5, characterized in that, The local statistical properties of the input signal are extracted through the adaptive wavelet decomposition module, including: The input signal is subjected to global average pooling to obtain a global context vector; The global context vector is input into the fully connected layer for nonlinear mapping to obtain the dynamic filter generation parameters. The low-pass filter weights and high-pass filter weights are generated based on the dynamic filter generation parameters; The input signal is subjected to a non-negative truncation constraint, and the input signal after the non-negative truncation constraint is convolved with the low-pass filter weights and the high-pass filter weights respectively. The absolute value activation is applied to the convolution result to obtain the approximation coefficients and detail coefficients.

7. The DAS spatiotemporal signal enhancement method for vehicle trajectory tracking according to claim 1, characterized in that, The detailed coefficients are input into the enhanced residual module of the fusion channel attention mechanism. Through depthwise separable convolution and SE channel attention mechanism, wavelet domain features are enhanced and weights are recalibrated to obtain enhanced detailed features, including: The detail coefficients are used as input features for the enhanced residual module; Spatial features are extracted from the input features by the depthwise separable convolution to obtain convolutional features; The channel dependency relationship of the convolutional features is modeled using the SE channel attention mechanism to obtain the channel weights. The convolutional features are recalibrated based on the channel weights to obtain recalibrated features. The input features and the recalibration features are subjected to residual fusion, and the residual fusion result is nonlinearly activated to obtain the enhanced detail features.

8. The DAS spatiotemporal signal enhancement method for vehicle trajectory tracking according to claim 1, characterized in that, In the decoding stage, the deep low-frequency components and the corresponding scale-enhanced detail features are aggregated at multiple scales to obtain a global aggregated representation, including: The deep low-frequency components are used as the initial reconstruction features in the decoding stage; The initial reconstructed features are upsampled to obtain upsampled reconstructed features; The upsampled reconstructed features and the enhanced detail features at the corresponding scale are concatenated along the channel dimension to obtain the aggregated features at the current scale; The aggregated features at the current scale are passed down to the shallow layers as the reconstructed features at the next scale until the global aggregated representation is obtained.

9. The DAS spatiotemporal signal enhancement method for vehicle trajectory tracking according to claim 1, characterized in that, The global aggregation representation is input into the parameterized reconstruction head for reconstruction mapping, and a non-negative projection operator is introduced at the end of the mapping of the parameterized reconstruction head to obtain a denoised reconstructed signal that conforms to the non-negative physical characteristics of DAS vibration signal energy, including: The global aggregate representation is input into the parameterized reconstruction head; The global aggregate representation is subjected to channel compression and temporal waveform reconstruction through multi-layer convolutional mapping in the parameterized reconstruction head to obtain the initial reconstructed signal. At the end of the mapping of the parameterized reconstruction head, the negative components in the initial reconstructed signal are filtered out by the non-negative projection operator; The initial reconstructed signal after filtering out negative components is used as the denoised reconstructed signal.

10. The DAS spatiotemporal signal enhancement method for vehicle trajectory tracking according to claim 3, characterized in that, The parameterized mapping function is optimized based on the reconstruction fidelity term and the wavelet basis orthogonal regularization term to form the reconstruction mapping, including: Construct a hybrid dataset consisting of relatively clean real signals and simulated synthetic data; Generating true samples of vehicle vibration signals based on a mass spring damping dynamic model of the vehicle-road coupled system; The true value sample of the vehicle vibration signal is decomposed into transient impact component and quasi-periodic oscillation component; Construct a heterogeneous noise environment that includes colored noise and Rayleigh distributed multiplicative noise; The true value sample of the vehicle vibration signal and the heterogeneous noise environment are superimposed to form the vibration observation signal of the training vehicle. The parameterized mapping function is optimized using the vibration observation signal of the training vehicle, so that the optimized parameterized mapping function forms the reconstructed mapping.

Images