DAS-based traffic monitoring technology system and signal processing method

Abstract

The invention relates to a vehicle-type-based traffic flow monitoring DAS system and a signal processing method, and belongs to the technical field of intelligent traffic monitoring. A DAS system for monitoring traffic flow based on different vehicle types is characterized by comprising a sensing module based on a phase-sensitive optical time domain reflection technology and a self-adaptive signal processing module, according to the system, firstly, road vibration data are collected by using an optical fiber distributed acoustic sensing technology, vehicle vibration characteristics are selectively enhanced and background interference is suppressed in a key frequency band of 10-80 Hz through DAS-PCSE based on enhanced ultra-wavelet transform, and high-stability input characterization is generated; then, the preprocessed data are input into a lightweight DAS Trajectory Net, the network takes an encoder-decoder as a skeleton and introduces an adaptive feature enhancement module, and accurate recognition of a vehicle track is achieved; and finally, through a DAS Veple Classifier Net, a double-current signal processing module and lightweight convolution are utilized to efficiently fuse spatio-temporal information, and vehicle type discrimination is completed. According to the invention, the influence of low signal-to-noise ratio and complex background disturbance can be effectively overcome, and wide-area and all-weather vehicle-type-based traffic flow and vehicle speed monitoring can be realized with low calculation cost.

Description

Technical Field

[0001] This invention relates to a traffic monitoring technology based on DAS and a signal processing method, belonging to the field of traffic monitoring technology. Background Technology

[0002] With deepening urbanization and the expansion of road networks, Intelligent Traffic Systems (ITS) are placing higher demands on traffic condition perception with high spatiotemporal resolution, long-distance coverage, and low cost. In recent years, a series of explorations have been undertaken regarding DAS traffic perception, but key bottlenecks remain. In terms of data preprocessing, the raw DAS signal is superimposed with environmental vibrations, system noise, and scattering noise, severely obscuring effective features induced by vehicles. Traditional methods often employ regularized filtering or wavelet thresholding for denoising, such as the cascaded use of Hampel outlier detection and bandpass filtering, as well as improved wavelet thresholding strategies. These methods can improve SNR in specific scenarios, but they typically rely on fixed parameters or a fixed basis, and their adaptability and discriminativeness remain limited when facing non-stationary and multi-source noise. In terms of feature representation and task modeling, DAS data naturally contains two complementary forms: spatiotemporal "waterfall plots" with strong spatial geometric structure, and local time-frequency signals containing fine resonance details. Existing image-based methods for vehicle type differentiation based on time-frequency differences and frequency band energy, and for trajectory line identification using Hough transform to estimate vehicle speed and traffic flow, are prone to degradation in weak morphology or multi-vehicle interference scenarios. Physical modeling methods such as beamforming and velocity superposition are also quite sensitive to parameters and signal-to-noise conditions. Although the introduction of deep learning has significantly improved the plasticity of DAS data, such as attention CNNs for denoising and enhancement, self-supervised U-Net architecture for spatiotemporal deconvolution and resolution enhancement, and speed estimation based on deep detection and regression networks (such as SpeedNet) and vehicle speed estimation based on YOLOv8, the paradigm of single end-to-end networks often ignores the task heterogeneity of "trajectory recognition (partially spatial domain)" and "vehicle classification (partially time-frequency domain)," increasing the optimization difficulty and limiting the targeted modeling of cross-domain features. At the same time, the computational and storage overhead of deep models poses a challenge to real-time deployment at the edge, and research on lightweight design and engineering usability still needs to be strengthened. Summary of the Invention

[0003] To address the problems existing in the prior art, a traffic monitoring technology based on DAS and a signal processing method are provided.

[0004] The present invention solves the above-mentioned technical problems through the following technical solution: A traffic monitoring technology based on DAS includes a phase-sensitive optical time-domain reflectostat, which is characterized by... In the optical time-domain reflectograph, continuous light emitted from a narrow-linewidth laser is modulated into nanosecond-level optical pulses by an acousto-optic modulator. These pulses are amplified by an erbium-doped fiber amplifier and then injected into sensing optical fibers laid along a road via a circulator. As the optical pulses propagate forward within the fiber, they interact with inherent Rayleigh scattering points in the fiber, generating weak backscattered Rayleigh light. This backscattered light carries information from points along its path and returns along the original route, again being guided by the circulator to a coherent detection module. In the coherent detection module, the returned signal light is mixed with the local oscillator light from the same laser and then converted into an electrical signal by a balanced photodetector. Finally, through high-speed analog-to-digital conversion and digital signal processing, the phase changes caused by external disturbances at each fiber location are demodulated.

[0005] Based on the above technical solution, this application further improves and refines the above technical solution as follows: A signal preprocessing method for the DAS of the above system is characterized by the following specific process. First, a 10–80 Hz bandpass filter is applied to the original spatiotemporal data to suppress low-frequency drift and high-frequency electronic noise. At the same time, the main vibration components during vehicle movement are preserved, making vehicle events more prominent in the time domain. Subsequently, an enhanced super wavelet transform is introduced. Based on the high time-frequency resolution characterization, the key frequency band signals and transient responses are specifically enhanced through frequency band selective weighting and phase consistency reconstruction, thereby improving the contrast between vehicle signals and background noise. Finally, the envelope of the enhanced signal is extracted using Hilbert transform. E ( t,x This demodulates the high-frequency carrier wave into a stable instantaneous energy curve.

[0006] Furthermore, the enhanced super wavelet transform achieves precise enhancement of key frequency bands of DAS vehicle signals through a series of steps including multi-order super wavelet transform, dynamic weighting, and phase reconstruction. The specific process is as follows: ① Multi-order Ultrawavelet Transform: The first step is to perform a multi-order ultrawavelet transform on the bandpass-filtered signal, corresponding to the first step in the workflow on the right. Unlike traditional wavelet transforms that use a fixed time-frequency window, ultrawavelet transforms achieve higher frequency resolution in the low-frequency band by combining wavelet basis functions of different orders (the order range is set to ord=(3,15) in this application). For each frequency point within the target frequency band (10-80Hz)... The algorithm will calculate the wavelet transform results of multiple orders k: Number of cycles =5, The total order is . For the sampling frequency, this process generates, as follows: Figure 3 The “Superlet Magnitude Spectrogram” shown in (A) reveals the distribution of signal energy in the time and frequency dimensions. It can be seen that the energy is mainly concentrated in 40–60 Hz when the vehicle passes by. ② Dynamic Weighting and Top-3 Selection Strategy: To further highlight key frequency bands, we defined an exponentially decaying weighting function centered on the target frequency band. The formula is as follows: in, β The weighting function represents the decay rate of the weights, where K represents the total number of frequency components. The weighting function is as follows: Figure 3 (B) As shown by the blue curve, the weight is the largest at the center frequency and gradually decreases towards both sides; Therefore, for each time point, the top-3 frequency components with the highest normalized amplitude are selected, and an exponentially decaying weight is applied with the target frequency band as the center to obtain the weighted amplitude spectrum Â(f,t). k =3). Calculate the weighted time-domain signal as follows: Figure 3 (B) As shown in the "Top-3 dynamic weighted sum" at the bottom. This strategy amplifies the contribution of key frequency bands by dynamically selecting and adapting to potential spectral differences among different vehicles; ③ Phase Smoothing and Reconstruction Algorithm: To enhance the signal while maintaining its physical authenticity and phase continuity, we designed this algorithm. The algorithm normalizes the spectrum, and the calculation formula is as follows: in, Represents the time spectrum of the super wavelet. Then, amplitude-weighted vector summation is performed on the complex spectral components within the main frequency band: in The dominant frequency index, representing the frequency with the highest average energy, is used to determine the dominant frequency component of the signal. T represents the total number of sampling points, thus obtaining a smooth, weighted phase determined by the dominant frequency. : like Figure 3As shown in (C), the original phases of multiple sub-bands (top figure, Sub-band phases) have varying linearity. After amplitude-weighted vector summation, a smooth phase with excellent linearity is obtained (bottom figure, Amplitude-weighted vector-sum phase). This is combined with the weighted amplitude spectrum obtained from the preceding steps. With smoothed phase The final enhanced signal is generated through complex reconstruction: like Figure 3 As shown in (D), the reconstructed signal has a clearer and more focused vibration waveform when the vehicle passes by compared to the original signal (Raw), noise is effectively suppressed, and the key instantaneous amplitude characteristics are significantly enhanced.

[0007] Furthermore, the envelope extraction module, This process is achieved using the Hilbert Transform, which can effectively demodulate a high-frequency oscillating carrier signal into a low-frequency envelope signal E(t,x) that characterizes its instantaneous energy change; Subsequently, for the two networks with different functions, DAS Trajectory Net and DAS VehicleClassifier Net, specialized samples adapted to their respective tasks were constructed: ① Sample construction of DAS Trajectory Net: The low-frequency envelope signal E(t,x) is standardized and used as the network input; the real trajectory mask is obtained through video-assisted manual annotation. The spatiotemporal location information of this mask can realize sample alignment and window selection, providing priors for trajectory recognition training and spatiotemporal constraints for subsequent vehicle signal classification; ② Sample Construction of DAS Vehicle Classifier Net: Based on the spatiotemporal prior of the above-mentioned labeled mask, a mapping is established between the mask pixel coordinates and the row and column indices of the envelope matrix E(t,x) to achieve precise positioning of trajectory points in time and space channels. On each spatial channel covered by the trajectory, the equal-weighted temporal centroid is calculated according to the temporal distribution of the trajectory points. Based on this, a fixed time window of 2 seconds (4096 sampling points) is extracted. Zero padding is applied to the out-of-bounds parts to eliminate boundary effects and unify the segment length. Finally, a feature sequence corresponding to a single vehicle, with normalized length and cross-channel alignment, is obtained as the standard input of the vehicle classification network.

[0008] Furthermore, the vehicle trajectory recognition (DAS Trajectory Net) includes, The vehicle trajectory recognition module—DAS Trajectory Net—has the following network architecture: Figure 5 As shown, this network adopts an encoder-decoder architecture and achieves accurate segmentation and extraction of vehicle trajectories in complex traffic environments through the lightweight design of the Adaptive Feature Enhancement (AFE) module. The specific structure of this module is as follows. ① Encoder Each level of the encoder adopts the DAS Trajectory Encoder structure, which includes standard convolution, batch normalization, ReLU activation, LCoV lightweight convolution, and max pooling operations, and can be represented as: ② Bottleneck layer To maintain a lightweight design while enhancing expressive power, the network introduces an Adaptive Feature Enhancement (AFE) module as a unified enhancement unit. For example... Figure 5 As shown in the upper right corner, the AFE module adopts an inverted residual architecture. It first expands the feature channel dimension through point convolution to improve the nonlinear representation capability, then uses depthwise separable convolution to efficiently extract spatial correlation, and finally uses point convolution projection to regress to a lower dimension.

[0009] The bottleneck layer consists of two consecutive AFE modules, enabling deep enhancement of high-level semantic features.

[0010] The dual AFE design can enhance the focus on global trajectory information in deep networks. By cascading adaptive feature enhancement operations, it captures the global semantic features of vehicle trajectory, laying the foundation for subsequent accurate pixel-level trajectory reconstruction.

[0011] ③ Decoder The decoder employs a DAS Trajectory Decoder structure, organically combining MaxUnpool upsampling, feature fusion, skip connection fusion, and AFE enhancement. The computational flow of the decoding fusion mechanism is as follows: Perform feature stitching and AFE processing: A global context enhancement mechanism is introduced to extract a one-dimensional feature vector through global average pooling: in and These represent the height and width of the feature map, respectively. One-dimensional convolution is used to exchange information between channels, strengthening the connections between global features without significantly increasing computational burden. in This is a one-dimensional convolution operation. This uses the Sigmoid activation function. Feature fusion is ultimately achieved through global feature weighting and residual connections. in Representing the Hard code product operation, this design ensures that the model can not only utilize local detailed features during upsampling, but also fully consider the contextual information of the global trajectory distribution, thereby achieving high-precision trajectory reconstruction.

[0012] Furthermore, the DAS Vehicle Classifier Net (Vehicle Type Recognition Network) ①Light Conv This application proposes a lightweight convolutional (LConv) module specifically for DAS vehicle signal classification tasks, based on a "partial activation, global preservation" feature processing strategy. This module first performs segmentation along the channel dimension: in, Represents the active channels participating in convolution calculations. This indicates a hold-alive channel using identity mapping.

[0013] The forward propagation process of LConv can be mathematically expressed as: Convolution operation Different convolutional kernel sizes can be selected according to specific task requirements. In the DualStream Signal Block, a 1×7 asymmetric convolutional kernel is used in the temporal domain to capture temporal dependencies, while a 7×1 convolutional kernel is used in the spatial domain branch to extract spatially relevant features.

[0014] ②Dual Stream Signal Block The dual-stream signal processing module is the core of the vehicle classification network. For example... Figure 6As shown, the input features are divided into two parallel branches by ChannelSplit. The temporal branch uses LConv 1×7 to capture transient vibration patterns when a vehicle passes by, while the spatial branch uses LConv 7×1 to focus on characterizing the spatial distribution features propagating along the optical fiber. The former focuses on short-term dynamic changes, while the latter characterizes the cooperative response and diffusion trend of adjacent locations. Then, the outputs of the two branches are fused through a Concat operation. The fused result is then subjected to 1×1 convolution to complete channel information interaction and dimensional reconstruction, achieving adaptive combination of spatiotemporal features using learnable weights. Finally, with the help of residual connections, the fused features are added element-wise to the original input, allowing the model to retain its memory of basic patterns while learning complex spatiotemporal relationships.

[0015] This application has the following advantages: I. Adaptive Signal Processing The key problem this invention addresses is that the effective features of vehicles are severely obscured by various noises, thus affecting the accuracy of subsequent trajectory recognition and vehicle classification. Correspondingly, adaptive signal processing based on superwavelet transform can specifically enhance the features of the core frequency band of vehicle vibration, effectively suppressing noise and forming standardized, high-contrast input data, enabling subsequent models to more clearly capture vehicle trajectories and type differences. This advantage stems from the proposed DAS-PCSE (Phase-Consistent Superlet Enhancement with Envelope Demodulation) preprocessing link. This link achieves signal optimization through three levels of processing: first, 10–80Hz bandpass filtering suppresses low-frequency drift and high-frequency electronic noise, focusing on the energy of the target frequency band; then, enhanced superwavelet transform (EST), combined with multi-order wavelet bases, dynamic frequency weighting, and phase reconstruction, strengthens the transient response and phase continuity of the key frequency band; finally, Hilbert transform is used to extract the envelope, demodulating the high-frequency carrier into a stable instantaneous energy curve, completely solving the problem of "weak features and strong noise" in the original signal.

[0016] II. Dual-model collaborative mechanism The primary problem this invention addresses is the fusion of features from multiple domains. Previous single-model end-to-end networks failed to distinguish between the tasks of "trajectory recognition (partially spatial domain)" and "vehicle classification (partially time-frequency domain)," leading to increased optimization difficulty and limiting targeted modeling of cross-domain features. In contrast, this invention offers advantages such as dual-model collaboration, efficient feature fusion, and an architecture that is easy to optimize. This advantage stems from the specialized design of the dual-model architecture: DAS Trajectory Net employs an encoder, decoder, and AFE module. The encoder uses lightweight LConv convolutions to extract spatial features and store downsampling indices. The bottleneck layer enhances global trajectory semantics through dual AFE modules. The decoder combines MaxUnpool upsampling and global context enhancement to meet the requirements of "spatial domain trajectory reconstruction." DAS Trajectory Net, specifically designed for trajectory recognition, can accurately capture the spatial motion path of vehicles. DAS VehicleClassifier Net is centered around a dual-stream signal processing module. Its time-domain branch (LConv 1×7) captures transient vehicle vibrations, while its spatial-domain branch (LConv 7×1) characterizes spatial diffusion features. Specifically designed for time-frequency domain vehicle classification, DAS VehicleClassifier Net can reliably distinguish between four vehicle types: sedans, medium-duty trucks, buses, and heavy-duty trucks. Its accuracy in both tasks surpasses most comparable models (such as U-Net and EfficientNet). The collaborative use of the two models addresses the task heterogeneity problem at its architectural root.

[0017] III. Lightweight Design and Edge Implementation Deployment Another problem this invention addresses is the resource-intensive nature of deep learning models. Traditional deep networks incur high computational and storage costs, making them difficult for edge devices to handle, and they lack cross-platform compatibility, failing to meet the engineering requirements of real-time traffic monitoring. To address this, the invention offers the dual advantages of high lightweight design and strong engineering usability. These advantages stem from lightweight module design and engineering validation. Firstly, regarding lightweight design, this method introduces structures such as Adaptive Feature Enhancement (AFE), Lightweight Convolution (LConv), and Dual-Stream Signaling (DSS) modules, significantly reducing FLOPs and parameter count while maintaining accuracy. LightConv reduces the number of parameters in traditional convolutions by approximately 75% through a "channel segmentation + partial convolution" strategy; the AFE module uses an inverse residual structure of "point convolution expansion - depthwise separable convolution - point convolution compression" to maintain feature representation with low computational overhead; the Dual-Stream Signaling (DSS) module replaces full-size convolutions with parallel branches, achieving a computational complexity only 1 / 4 that of traditional 7×7 convolutions. This design verifies the compatibility with multiple GPU platforms through testing and combines end-to-end visualization to ensure that the model can be deployed and monitored in real-world scenarios, thus solving the problem of "heavy and difficult-to-deploy models". Attached Figure Description

[0018] Figure 1 This is a structural diagram of a distributed acoustic sensing system. Figure 2 This is an adaptive signal preprocessing workflow based on ultrawavelet transform; Figure 3 Flowchart of enhanced superwavelet transform; Figure 4 A framework diagram for traffic flow monitoring; Figure 5 For DAS Trajectory Net; Figure 6 For Vehicle Classifier Net; Figure 7 This is the overall flowchart. Detailed Implementation

[0019] The following embodiments, in conjunction with the accompanying drawings, are merely for illustrating the technical solutions described in the claims and are not intended to limit the scope of protection of the claims.

[0020] I. Signal Acquisition Equipment The traffic monitoring technology employed in this invention is based on a distributed acoustic sensing (DAS) system, the core of which is a phase-sensitive optical time-domain reflectometer (Φ-OTDR). For example... Figure 1 As shown, continuous light emitted from a narrow-linewidth laser is modulated into nanosecond-level optical pulses by an acousto-optic modulator. These pulses are amplified by an erbium-doped fiber amplifier (EDFA) and then injected into sensing optical fibers laid along the road via a circulator. As the optical pulses propagate forward within the fiber, they interact with the inherent Rayleigh scattering points in the fiber, generating weak backscattered Rayleigh light. This scattered light carries information from various points along the path and returns along the same route, again being guided by the circulator to the coherent detection module. In this module, the returned signal light is mixed with the local oscillator light from the same laser and then converted into an electrical signal by a balanced photodetector (BPD). Finally, through high-speed analog-to-digital conversion (A / D) and digital signal processing, the phase changes caused by external disturbances (such as vehicle vibrations) at each fiber location are demodulated. It is this extremely high sensitivity to phase that enables the DAS system to accurately reconstruct the spatiotemporal dynamic distribution of vibrations over a range of hundreds of kilometers with meter-level spatial resolution and millisecond-level temporal resolution, providing a high-quality data foundation for subsequent intelligent analysis.

[0021] II. DAS Signal Preprocessing Flow To improve the usability of raw DAS vibration signals in traffic scenarios, this paper constructs a three-level data processing flow (as shown above). Figure 2 As shown): First, a 10–80 Hz bandpass filter is applied to the original spatiotemporal data to suppress low-frequency drift and high-frequency electronic noise, while retaining the main vibration components during vehicle motion, making vehicle events more prominent in the time domain. Then, an enhanced superlet transform (EST) is introduced. Based on the high time-frequency resolution representation, key frequency band signals and transient responses are specifically enhanced through frequency band selective weighting and phase-consistent reconstruction, improving the contrast between vehicle signals and background noise. Finally, a Hilbert transform is used to extract the envelope E(t,x) of the enhanced signal, demodulating the high-frequency carrier into a stable instantaneous energy curve, preserving the intensity evolution characteristics of the vehicle while significantly reducing the interference of morphological fluctuations on subsequent modeling. This three-stage processing realizes the transformation of the original weak DAS signal into a "highly discriminative, standardizable input," providing reliable data support for subsequent trajectory recognition and vehicle classification. The EST module and envelope extraction module in data preprocessing will be described in detail below: 1. Enhanced Ultrawavelet Transform Module The enhanced superlet transform is the core of the signal preprocessing workflow. For example... Figure 3 As shown, this algorithm achieves precise enhancement of key frequency bands of DAS vehicle signals through a series of steps including multi-order ultrawavelet transform, dynamic weighting, and phase reconstruction. The specific process is as follows: ① Multi-order Ultrawavelet Transform: The first step is to perform a multi-order ultrawavelet transform on the bandpass-filtered signal, corresponding to the first step in the workflow on the right. Unlike traditional wavelet transforms that use a fixed time-frequency window, ultrawavelet transforms achieve higher frequency resolution in the low-frequency band by combining wavelet basis functions of different orders (the order range is set to ord=(3,15) in this application). For each frequency point within the target frequency band (10-80Hz)... The algorithm will calculate the wavelet transform results of multiple orders k: Number of cycles =5, The total order is . The sampling frequency; Represents the frequency domain amplitude after super wavelet transform, ( ord 0 , ord 1 ） Represents the range of orders. CWT Represents wavelet transform, where x represents the signal after bandpass filtering., This process generated, as Figure 3 The “Superlet Magnitude Spectrogram” shown in (A) reveals the distribution of signal energy in the time and frequency dimensions. It can be seen that the energy is mainly concentrated in 40–60 Hz when the vehicle passes by.

[0022] ② Dynamic Weighting and Top-3 Selection Strategy: To further highlight key frequency bands, we defined an exponentially decaying weighting function centered on the target frequency band. The formula is as follows: in, β This indicates the decay rate of the weights. K Indicates the total number of frequency components. k Represents the current frequency channel index. j This represents the indices of all frequency channels used for normalized summation. The weighting function is as follows: Figure 3 (B) As shown by the blue curve, the weight is the largest at the center frequency and gradually decreases towards both sides.

[0023] Therefore, for each time point, the top-3 frequency components with the highest normalized amplitude are selected, and an exponentially decaying weight is applied centered on the target frequency band to obtain the weighted amplitude spectrum Â(f,t) (k=3). The weighted time-domain signal is calculated as follows: Figure 3 (B) As shown in the "Top-3 dynamic weighted sum" at the bottom. This strategy amplifies the contribution of key frequency bands and adapts to the possible spectral differences between different vehicles through dynamic selection.

[0024] ③ Phase Smoothing and Reconstruction Algorithm: To enhance the signal while maintaining its physical authenticity and phase continuity, we designed this algorithm. The algorithm normalizes the spectrum, and the calculation formula is as follows: in, M norm [ f,t [Represents the normalized spectrum] This represents the maximum amplitude of the frequency f over all times t. Represents the time spectrum of the super wavelet. Then, amplitude-weighted vector summation is performed on the complex spectral components within the main frequency band: in The dominant frequency index, which represents the frequency with the highest average energy, is used to determine the dominant frequency component of a signal. This represents the total number of sampling points, thus yielding a smooth, weighted phase determined by the dominant frequency. : like Figure 3 As shown in (C), the original phases of multiple sub-bands have varying linearity. After amplitude-weighted vector summation, a smooth phase with excellent linearity is obtained. This is combined with the weighted amplitude spectrum obtained from the preceding steps. With smoothed phase The final enhanced signal is generated through complex reconstruction: like Figure 3 As shown in (D), the reconstructed signal has a clearer and more focused vibration waveform when the vehicle passes by compared to the original signal (Raw), noise is effectively suppressed, and the key instantaneous amplitude characteristics are significantly enhanced.

[0025] 2. Envelope Extraction Module The spatiotemporal data matrix after enhanced ultrawavelet transform (EST) processing Perform instantaneous envelope extraction.

[0026] This application employs the Hilbert Transform to implement this process, which effectively demodulates a high-frequency oscillating carrier signal into a low-frequency envelope signal E(t,x) characterizing its instantaneous energy change. Subsequently, we construct specialized samples adapted to the tasks of two networks with different functions (DAS Trajectory Net and DAS Vehicle Classifier Net): ① Sample construction of DAS Trajectory Net: The low-frequency envelope signal E(t,x) is standardized and used as the network input; the real trajectory mask is obtained through video-assisted manual annotation. The spatiotemporal location information of this mask can realize sample alignment and window selection, providing priors for trajectory recognition training and spatiotemporal constraints for subsequent vehicle signal classification.

[0027] ② Sample Construction of DAS Vehicle Classifier Net: Based on the spatiotemporal prior of the above-mentioned labeled mask, a mapping is established between the mask pixel coordinates and the row and column indices of the envelope matrix E(t,x) to achieve precise positioning of trajectory points in time and space channels. On each spatial channel covered by the trajectory, the equal-weighted temporal centroid is calculated according to the temporal distribution of the trajectory points. Based on this, a fixed time window of 2 seconds (4096 sampling points) is extracted. Zero padding is applied to the out-of-bounds parts to eliminate boundary effects and unify the segment length. Finally, a feature sequence corresponding to a single vehicle, with normalized length and cross-channel alignment, is obtained as the standard input of the vehicle classification network.

[0028] III. Traffic Flow Detection Principle Framework This invention proposes a lightweight dual-model collaborative mechanism for DAS traffic monitoring, the overall process of which is as follows: Figure 4 As shown, this framework divides the complex monitoring task into two sub-tasks: vehicle trajectory recognition (DAS Trajectory Net) and vehicle type recognition (DAS Vehicle Classifier Net), and designs a specialized network architecture. The relevant modules of the two sub-tasks will be introduced one by one below.

[0029] 1. DAS Trajectory Net (Vehicle Trajectory Recognition Network) First, let's introduce the vehicle trajectory recognition module—DAS Trajectory Net, whose network architecture is as follows: Figure 5 As shown, the network adopts an encoder-decoder architecture and achieves accurate segmentation and extraction of vehicle trajectories in complex traffic environments through the lightweight design of the Adaptive Feature Enhancement (AFE) module. The following is a detailed introduction to this module.

[0030] ① Encoder Design Each level of the encoder adopts the DAS Trajectory Encoder structure, which includes standard convolution, batch normalization, ReLU activation, LCoV lightweight convolution, and max pooling operations, and can be represented as: ② Bottleneck layer design To maintain a lightweight design while enhancing expressive power, the network introduces an Adaptive Feature Enhancement (AFE) module as a unified enhancement unit. For example... Figure 5 As shown in the upper right corner, the AFE module adopts an inverted residual architecture. It first expands the feature channel dimension through point convolution to improve the nonlinear representation capability, then uses depthwise separable convolution to efficiently extract spatial correlation, and finally uses point convolution projection to regress to a lower dimension.

[0031] The bottleneck layer consists of two consecutive AFE modules, enabling deep enhancement of high-level semantic features.

[0032] in, Bottleneeck ( X The ) represents the feature map output after computation by the bottleneck layer. AFE 2 and AFE 1 represents the computation of the adaptive feature enhancement module. The dual AFE design can enhance the attention to global trajectory information in deep networks. By cascading adaptive feature enhancement operations, it captures the global semantic features of vehicle trajectory, laying the foundation for subsequent accurate pixel-level trajectory reconstruction.

[0033] ③ Decoder Design The decoder employs a DAS Trajectory Decoder structure, organically combining MaxUnpool upsampling, feature fusion, skip connection fusion, and AFE enhancement. The computational flow of the decoding fusion mechanism is as follows: Perform feature stitching and AFE processing: in, F represents the feature map passed from the previous layer module and the feature map transmitted through the skip link, respectively. concat The representative will The feature map obtained by concatenating along the channel dimension Concat [] represents a concatenation operation along the feature map channel dimension. F afe This represents the feature map calculated by the AFE module in the decoder.

[0034] A global context enhancement mechanism is introduced to extract a one-dimensional feature vector through global average pooling: in, F global This represents a one-dimensional feature vector calculated using global average pooling. and These represent the height and width of the feature map, respectively. One-dimensional convolution is used to exchange information between channels, strengthening the connections between global features without significantly increasing computational burden. in, W global Represents global eigenvalues. This is a one-dimensional convolution operation. This uses the Sigmoid activation function. Feature fusion is ultimately achieved through global feature weighting and residual connections. in, F outbal Represents the output feature map of each decoder. Representing the Hard code product operation, this design ensures that the model can not only utilize local detailed features during upsampling, but also fully consider the contextual information of the global trajectory distribution, thereby achieving high-precision trajectory reconstruction.

[0035] 2. DAS Vehicle Classifier Net ①Light Conv This application proposes a lightweight convolutional (LConv) module specifically for DAS vehicle signal classification tasks, based on a "partial activation, global preservation" feature processing strategy. This module first performs segmentation along the channel dimension: in Represents the active channels participating in convolution calculations. This indicates a hold-alive channel using identity mapping.

[0036] The forward propagation process of LConv can be mathematically expressed as: Convolution operation Different convolutional kernel sizes can be selected according to specific task requirements. In the DualStream Signal Block, a 1×7 asymmetric convolutional kernel is used in the temporal domain to capture temporal dependencies, while a 7×1 convolutional kernel is used in the spatial domain branch to extract spatially relevant features.

[0037] This represents the feature map calculated using a 1×7 asymmetric convolution kernel. This represents the feature map calculated using a 7×1 asymmetric convolution kernel. Compared to traditional full-channel convolution operations, the LCoV module significantly reduces computational complexity while maintaining feature representation capabilities. The traditional convolution has a parameter count of... The number of parameters for LConv is only K 2 The product representing the kernel size achieves approximately 75% parameter reduction. Simultaneously, the channels preserved through the identity mapping ensure efficient gradient propagation, preventing information loss in deep networks.

[0038] ②Dual Stream Signal Block The dual-stream signal processing module is the core of the vehicle classification network. For example... Figure 6 As shown, the input features are divided into two parallel branches by ChannelSplit. The temporal branch uses LConv 1×7 to capture transient vibration patterns when a vehicle passes by, while the spatial branch uses LConv 7×1 to focus on characterizing the spatial distribution features propagating along the optical fiber. The former focuses on short-term dynamic changes, while the latter characterizes the cooperative response and diffusion trend of adjacent locations. Then, the outputs of the two branches are fused through a Concat operation. The fused result is then subjected to 1×1 convolution to complete channel information interaction and dimensional reconstruction, achieving adaptive combination of spatiotemporal features using learnable weights. Finally, with the help of residual connections, the fused features are added element-wise to the original input, allowing the model to retain its memory of basic patterns while learning complex spatiotemporal relationships.

[0039] Compared to the computational complexity of traditional full-channel 7×7 convolution... The computational complexity of the dual-stream signal processing module is only [missing information]. Meanwhile, parallel computing further improves inference efficiency.

[0040] As shown in Figure 7, this embodiment demonstrates a traffic flow monitoring system based on distributed acoustic sensing (DAS) and its specific implementation path, which mainly consists of two consecutive stages: data acquisition at the physical layer and data processing and analysis at the logical layer.

[0041] In the physical layer data acquisition phase, the hardware sensing environment is first constructed. Standard single-mode communication optical fibers are linearly laid along the roadside. These fibers act as passive sensors, detecting mechanical vibrations caused by vehicles traveling on the road. One end of the fiber is connected to a DAS signal acquisition device, which sends laser pulses into the fiber and receives backscattered Rayleigh light carrying vibration information. Phase vibration data at various locations along the fiber is acquired through photoelectric conversion and demodulation. The DAS acquisition device transmits the demodulated raw data to a personal computer or server via a high-speed interface for further processing. Furthermore, to build a training dataset or verify system accuracy, a video recording device is also installed on the roadside as a ground truth reference to synchronously record actual vehicle traffic.

[0042] In the data processing and analysis phase of the logic layer, the computing unit performs pipelined processing on the acquired DAS signals. The system first receives the raw DAS vibration signal mixed with a large amount of noise and initiates the signal preprocessing process: first, it removes low-frequency environmental interference and high-frequency system noise through bandpass filtering; then, it uses enhanced ultrawavelet transform (EST) for time-frequency analysis to focus on the core energy frequency band of vehicle vibration; finally, it converts the high-frequency oscillation signal into a low-frequency envelope signal reflecting energy intensity through envelope extraction, thereby converting the original blurred waterfall diagram into a black-and-white energy diagram with a clean background and clear trajectory.

[0043] Subsequently, the preprocessed data enters the deep learning analysis stage. DAS Trajectory Net first receives the preprocessed spatiotemporal image, uses deep learning algorithms for semantic segmentation, automatically identifies and extracts the diagonal trajectories representing vehicle motion, and generates a binary trajectory mask, thus solving the vehicle localization and counting problem. Based on this mask, the system accurately extracts vibration data segments belonging to specific vehicles from the original signal and inputs them into DAS Vehicle ClassifierNet. This classification network analyzes the temporal waveform characteristics and spatial diffusion characteristics of the vibration signal to determine the vehicle type, classifying vehicles into categories such as sedans, medium-sized trucks, buses, and heavy-duty trucks.

[0044] Ultimately, the system integrates trajectory and classification information to output monitoring results. On the monitoring interface, the system draws and visualizes vehicle trajectories in real time, calculating vehicle speed and arrival time based on trajectory slope and coordinates. Simultaneously, the system outputs statistical distribution maps of different vehicle types based on the classification results, achieving accurate statistics and display of traffic flow by vehicle type.

[0045] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A traffic monitoring technology based on DAS system, comprising a phase-sensitive optical time-domain reflectostat, characterized in that, In the optical time domain reflector, the continuous light emitted by the narrow linewidth laser is modulated into nanosecond-level optical pulses by an acousto-optic modulator. The optical pulses are amplified by an erbium-doped fiber amplifier and then injected into the sensing fiber laid along the road through a circulator. When a light pulse propagates forward within an optical fiber, it interacts with the inherent Rayleigh scattering points in the fiber, generating weak backscattered Rayleigh light. This backscattered light carries information from each point along its path and returns along the same route, then is guided again by a circulator to the coherent detection module. In the coherent detection module, the returned signal light is mixed with the local oscillator light from the same laser, and then converted into an electrical signal by a balanced photodetector. Finally, through high-speed analog-to-digital conversion and digital signal processing, the phase changes caused by external disturbances at each fiber location are demodulated.

2. A signal preprocessing method for the DAS of the system according to claim 1, characterized in that, The specific process is as follows: First, a 10–80 Hz bandpass filter is applied to the raw signal acquired by the DAS system to suppress low-frequency drift and high-frequency electronic noise. At the same time, the main vibration components during vehicle movement are preserved, making vehicle events more prominent in the time domain. Subsequently, an enhanced super wavelet transform is introduced. Based on the high time-frequency resolution characterization, the key frequency band signals and transient responses are specifically enhanced through frequency band selective weighting and phase consistency reconstruction, thereby improving the contrast between vehicle signals and background noise. Finally, the envelope extraction module uses Hilbert transform to extract the envelope of the enhanced signal. E ( t,x This demodulates the high-frequency carrier wave into a stable instantaneous energy curve.

3. The signal preprocessing method for DAS according to claim 2, characterized in that, The enhanced super wavelet transform achieves precise enhancement of key frequency bands of DAS vehicle signals through multi-order super wavelet transform, dynamic weighting, and phase reconstruction. The specific process is as follows: ① Multi-order ultrawavelet transform: First, a multi-order ultrawavelet transform is performed on the bandpass filtered signal. By combining wavelet basis functions of different orders, higher frequency resolution can be achieved in the low-frequency band; for each frequency point within the target frequency band... The algorithm will calculate the wavelet transform results of multiple orders k: S L T f i = 1 N o r d ∑ k = o r d 0 o r d 1 C W T x f i c k f i / f r e f 0.5 ， Number of cycles =5, The total order is . The sampling frequency; ② Dynamic Weighting and Top-3 Selection Strategy: To further highlight key frequency bands, an exponentially decaying weighting function centered on the target frequency band is defined. The formula is as follows: ， Where β represents the decay rate of the weight, K Indicates the total number of frequency components; ③ Phase smoothing and reconstruction algorithm: In order to enhance the signal while ensuring its physical authenticity and phase continuity, the algorithm normalizes the spectrum, and the calculation formula is as follows: ， in, M n o r m f t This represents the signal after phase smoothing. S f t Represents the time spectrum of the ultrawavelet. Then, amplitude-weighted vector summation is performed on the complex spectral components within the main frequency band: ， in, The dominant frequency index, which represents the frequency with the highest average energy, is used to determine the dominant frequency component of a signal. T This represents the total number of sampling points, thus yielding a smooth, weighted phase determined by the dominant frequency. φ w e i g h t e d t : , The weighted amplitude spectrum obtained by combining the above steps M n o r m f t ⋅ w k With smoothed phase φ w e i g h t e d t The final enhanced signal is generated through complex reconstruction: 。 4. The signal preprocessing method for DAS according to claim 2, characterized in that, The envelope extraction module, This process is achieved using the Hilbert transform, which can effectively demodulate a high-frequency oscillating carrier signal into a low-frequency envelope signal E(t,x) that characterizes its instantaneous energy change; Subsequently, for the two networks with different functionalities, DAS Trajectory Net and DAS Vehicle ClassifierNet, specialized samples adapted to their respective tasks were constructed: ① Sample construction of DAS Trajectory Net: The low-frequency envelope signal E(t,x) is standardized and used as the network input; the real trajectory mask is obtained by video-assisted manual annotation; the spatiotemporal location information of the mask can realize sample alignment and window selection, which not only provides priors for trajectory recognition training, but also provides spatiotemporal constraints for subsequent vehicle signal classification. ② Sample construction of DAS Vehicle Classifier Net: Based on the spatiotemporal prior of the above-mentioned labeled mask, a mapping between the mask pixel coordinates and the row and column indices of the envelope matrix E(t,x) is established to achieve accurate positioning of trajectory points in time and space channels; on each spatial channel covered by the trajectory, the equal-weighted time centroid is calculated according to the time distribution of the trajectory points, and a fixed time window of 2 seconds is extracted based on this; zero padding is applied to the out-of-bounds parts to eliminate boundary effects and unify the segment length, finally obtaining a feature sequence corresponding to a single vehicle, with normalized length and cross-channel alignment, which serves as the standard input of the vehicle classification network.

5. The signal preprocessing method for DAS according to claim 4, characterized in that, The DAS TrajectoryNet network adopts an encoder-decoder architecture and achieves accurate segmentation and extraction of vehicle trajectories in complex traffic environments through a lightweight design of an adaptive feature enhancement module. The specific structure of this module is as follows. ① Encoder Each level of the encoder adopts the DAS Trajectory Encoder structure, which includes standard convolution, batch normalization, ReLU activation, LCoV lightweight convolution, and max pooling operations, and can be represented as: , ② Bottleneck layer To maintain lightweight design while enhancing expressive power, the network introduces an adaptive feature enhancement module as a unified enhancement unit. The AFE module adopts an inverted residual architecture, first expanding the feature channel dimension through point convolution to improve nonlinear representation capability, then efficiently extracting spatial correlation through depthwise separable convolution, and finally projecting back to a lower dimension through point convolution. The bottleneck layer consists of two consecutive AFE modules, enabling deep enhancement of high-level semantic features; , The dual AFE design can enhance the focus on global trajectory information in deep networks and capture the global semantic features of vehicle trajectory through cascaded adaptive feature enhancement operations, laying the foundation for subsequent accurate pixel-level trajectory reconstruction. ③ Decoder The decoder adopts the DAS Trajectory Decoder structure, which organically combines MaxUnpool upsampling, feature fusion, skip connection fusion, and AFE enhancement; the computation flow of the decoding fusion mechanism is as follows: Perform feature stitching and AFE processing: ， A global context enhancement mechanism is introduced to extract a one-dimensional feature vector through global average pooling: ， in H and W The height and width of the feature maps, respectively, are processed by one-dimensional convolution to exchange information between channels, thereby strengthening the connections between global features without significantly increasing the computational burden. ， in, This is a one-dimensional convolution operation. The activation function is Sigmoid; feature fusion is finally achieved through global feature weighting and residual connections. ， in, This represents the Hardy code product operation.

6. The signal preprocessing method for DAS according to claim 4, characterized in that, The DAS VehicleClassifier Net: ①Light Conv Based on the feature processing strategy of "partial activation, global preservation", the first step is to perform segmentation along the channel dimension: ， in, Represents the active channels participating in convolution calculations. This indicates a hold-alive channel using identity mapping; The forward propagation process of LConv can be mathematically expressed as: , Among them, convolution operation Different kernel sizes can be selected according to specific task requirements; in the DualStream Signal Block, a 1×7 asymmetric kernel is used in the temporal domain to capture temporal dependencies, while a 7×1 kernel is used in the spatial domain branch to extract spatially relevant features. ， ， ②Lual Stream Signal Block, The input features are divided into two parallel branches by Channel Split. The temporal branch uses LConv 1×7 to capture transient vibration patterns when a vehicle passes by, while the spatial branch uses LConv 7×1 to focus on characterizing the spatial distribution features propagating along the optical fiber. The former focuses on short-term dynamic changes, while the latter characterizes the cooperative response and diffusion trend of adjacent locations. Then, the outputs of the two branches are fused through a Concat operation. The fused result is then subjected to 1×1 convolution to complete channel information interaction and dimension reconstruction, achieving adaptive combination of spatiotemporal features. Learnable weights are used to achieve adaptive combination of spatiotemporal features. Finally, with the help of residual connections, the fused features are added element-wise to the original input, allowing the model to retain the memory of basic patterns while learning complex spatiotemporal correlations.

Images