Fiber traffic signal data enhancement method based on masking autoencoder
By implicitly extracting the global topology of fiber optic signals through a masked autoencoder, high-fidelity twin samples are generated, solving the problems of data scarcity and signal breakage in fiber optic traffic monitoring and improving the robustness and detection accuracy of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2026-04-23
- Publication Date
- 2026-07-21
AI Technical Summary
Existing distributed fiber optic traffic monitoring technologies face challenges such as the scarcity of high-quality, fine-grained labeled data, signal breakage caused by real and complex environments, and difficulties in reusing self-supervised learning labels, making it difficult to generate high-fidelity, noise-resistant, and robust twin training samples.
A masked autoencoder-based approach is adopted, which implicitly extracts the global topology of vehicle vibration signals through high-ratio random masking and asymmetric heavyweight encoders. A lightweight decoder is then used for pixel-level reconstruction to generate twin samples with high-precision labels.
Without increasing the cost of manual annotation, we achieved zero-cost expansion of high-quality twin samples, maintained the physical and semantic consistency of the signal, and improved the robustness and generalization ability of downstream models.
Smart Images
Figure CN122435552A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for enhancing fiber optic traffic signal data based on a masked autoencoder, belonging to the field of intelligent transportation and fiber optic sensing data processing technology. Background Technology
[0002] With the rapid development of Intelligent Transportation Systems (ITS), Distributed Fiber Optic Sensing (DFOS) technology has demonstrated great potential in traffic flow management due to its advantages of large-scale, continuous, and non-destructive monitoring. By utilizing standard roadside communication or sensing optical cables, DFOS technology can transform infrastructure into a continuous virtual sensor array. After the vibration response caused by vehicles traveling on the road surface is demodulated, a two-dimensional vibration signal matrix containing rich spatiotemporal features (often called a "waterfall plot") is formed. Deep learning models can then treat this matrix as a visual domain image, extract the vehicle trajectory ridges, and thus achieve vehicle detection and instance segmentation.
[0003] However, in practical engineering deployments and large-scale networking applications, deep learning-based fiber optic traffic monitoring technology faces the following severe bottlenecks:
[0004] First, high-quality, fine-grained labeled data is extremely scarce, and the pre-training paradigm has limitations. The performance of deep learning perception models is highly dependent on large-scale labeled datasets. However, in the field of fiber optic sensing, pixel-level high-precision labeling for vehicle instances is extremely costly. To alleviate this problem, existing research often adopts a pre-training-fine-tuning paradigm (i.e., loading pre-trained weights based on general visual datasets such as ImageNet), but there is a huge semantic gap between general visual images and fiber optic spatiotemporal waterfall plots (which have strong stripe structures and long-range spatiotemporal continuity). Furthermore, simple weight initialization cannot fundamentally expand the training sample size in the fiber optic domain, and the model is still prone to overfitting when faced with small sample (Few-shot) conditions.
[0005] Second, the complex real-world environment leads to signal breakage, making it difficult for traditional augmentation methods to maintain global topology consistency. In real road networks, fiber optic signals are highly susceptible to strong interference from transient environmental noise (such as construction noise and wind noise), or local breaks and blockages in the effective vibration signal due to vehicle intersections and poor coupling. Traditional image-level data augmentation methods (such as random geometric transformations, color jittering, or conventional generative adversarial networks (GANs) often only introduce random perturbations at the local pixel level. They cannot enable the model to truly learn the underlying global topology and physical continuity of vehicle vibration signals, and the generated samples often exhibit physical trajectory distortion or structural collapse, making them unsuitable as high-quality auxiliary training samples.
[0006] Third, the potential of self-supervised learning in fiber optic data augmentation has not been fully explored, and label reuse remains difficult. In recent years, self-supervised learning based on Masked Image Modeling (MIM) has demonstrated powerful feature representation capabilities in general vision. However, in the field of distributed fiber optic sensing, masked autoencoders (MAEs) are mostly used only as implicit feature extractors. How to leverage their powerful context-aware repair and reconstruction capabilities to recover the spatiotemporal topology of fiber optic signals under extremely high masking rates (such as simulating severe signal loss), and thereby directly generate twin samples with the original high-precision labels to achieve zero-cost label reuse, remains a subject of little research.
[0007] Therefore, in the face of fiber optic traffic monitoring scenarios with complex interference, how to propose a data augmentation method that can implicitly extract global topological priors of signals and generate high-fidelity, noise-resistant, and robust twin training samples through self-supervised reconstruction without increasing any manual annotation costs has become a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0008] To address the problem that existing methods for enhancing distributed fiber optic traffic sensor data easily disrupt the global physical topology continuity of fiber optic signals, this invention provides a fiber optic traffic signal data enhancement method based on a masked autoencoder.
[0009] The present invention provides a fiber optic traffic signal data enhancement method based on a masked autoencoder, comprising:
[0010] The system acquires raw spatiotemporal vibration signals collected by distributed fiber optic sensing devices during vehicle operation, generates a standardized two-dimensional spatiotemporal matrix for a single channel, and then configures the original set of fine-grained artificial annotations.
[0011] The standardized two-dimensional spatiotemporal matrix is divided into multiple regular and non-overlapping image blocks, and a high proportion of random masking is applied to obtain visible block sequences and masked block sequences.
[0012] The visible block sequence and classification label are input into a heavyweight encoder based on a self-attention mechanism for feature mapping to obtain a low-dimensional latent feature representation vector;
[0013] The low-dimensional latent feature representation vector and the masking markers corresponding to the masking block sequence are concatenated and combined according to their original spatial positions to obtain the concatenated complete sequence. Then, a lightweight decoder is used to optimize the concatenated complete sequence by minimizing the mean square error of the region corresponding to the masking block sequence. Combined with the contextual semantic information of the visible block sequence, pixel-level signal reconstruction is performed to obtain the reconstructed spatiotemporal signal image.
[0014] The original set of fine-grained artificial annotations is mapped onto the reconstructed spatiotemporal signal image to obtain twin-enhanced samples with high-precision labels.
[0015] According to the fiber optic traffic signal data enhancement method based on masked autoencoders of the present invention, the heavyweight encoder captures the inherent global topology and long-range spatiotemporal dependence of the original spatiotemporal vibration signal based on the visible block sequence to obtain a low-dimensional latent feature representation vector.
[0016] According to the fiber optic traffic signal data enhancement method based on masking autoencoder of the present invention, the lightweight decoder infers and completes the corresponding region of the masking block sequence based on the contextual semantic information of the visible block sequence to obtain the reconstructed spatiotemporal signal image.
[0017] According to the fiber optic traffic signal data augmentation method based on masked autoencoders of the present invention, the original spatiotemporal vibration signal is aggregated with the twin augmentation samples to construct an augmentation training set. :
[0018] ,
[0019] In the formula, N represents the number of original spatiotemporal vibration signals. For the i-th original spatiotemporal vibration signal, Let i be the set of original artificial fine-grained annotations corresponding to the i-th original spatiotemporal vibration signal. For the i-th reconstructed spatiotemporal signal image, The reconstruction mapping function is the masked autoencoder after convergence trained on large-scale unsupervised data. The masked autoencoder is a complete asymmetric network consisting of a heavyweight encoder and a lightweight decoder. Let be the random binary masking matrix corresponding to the random masking rate.
[0020] According to the fiber optic traffic signal data enhancement method based on a masking autoencoder of the present invention, the random masking rate is not less than 75%.
[0021] According to the fiber optic traffic signal data augmentation method based on masked autoencoders of the present invention, the augmentation training set... Used for training downstream multi-task deep learning perception models.
[0022] According to the fiber optic traffic signal data enhancement method based on masked autoencoder of the present invention, the multi-task deep learning perception model includes an instance segmentation model, an object detection model, and a semantic segmentation network model.
[0023] According to the fiber optic traffic signal data enhancement method based on masked autoencoder of the present invention, the distributed fiber optic sensing device acquires data including DAS, DVS and OFDR.
[0024] According to the fiber optic traffic signal data enhancement method based on masked autoencoder of the present invention, the original spatiotemporal vibration signal is preprocessed by denoising and downsampling to generate a single-channel standardized two-dimensional spatiotemporal matrix.
[0025] According to the fiber optic traffic signal data enhancement method based on masking autoencoder of the present invention, the original artificial fine-grained annotation set includes instance segmentation mask and bounding box (Bbox).
[0026] The beneficial effects of this invention: The method of this invention is applied to edge computing nodes or cloud servers of distributed fiber optic sensing devices, and is based on masked autoencoders and self-supervised reconstruction. It cleverly utilizes the context-aware and structural repair capabilities of asymmetric masked autoencoders to implicitly extract long-range topological priors of vehicle vibration signals under extremely high masking rates. Furthermore, while strictly maintaining the physical semantic space layout, it achieves zero-cost expansion and adaptive denoising of high-quality labeled twin samples. Specifically:
[0027] An innovative, lossless annotation and reuse mechanism breaks through the bottleneck of data scarcity at zero cost: This invention abandons the traditional data synthesis method that requires manual intervention, cleverly utilizing the contextual semantic recovery capability of a masked autoencoder. Because the reconstructed image strictly preserves the spatial location, geometric shape, and category attributes of the original vehicle vibration trajectory, the system can directly map the original high-precision instance segmentation mask and bounding box onto the reconstructed image. This process achieves a significant increase in high-quality, labeled training samples without adding any additional manual annotation costs, fundamentally alleviating the pain point of extremely scarce labeled data in the field of fiber optic traffic perception.
[0028] Implicit extraction of global topological priors maintains strict physical semantic consistency: Unlike conventional image enhancement methods (such as random geometric transformations or direct pixel noise addition) that easily disrupt the unique long-range continuity of fiber optic signals, this invention introduces a high proportion of random masking and an asymmetric heavyweight encoder to force the network to move beyond local texture details and instead learn the inherent global spatiotemporal dependencies of vehicle vibration signals. Even with an extremely high rate of local information loss, the model can still infer and complete the broken signal, ensuring that the generated "twin" samples not only avoid the topological distortions common in conventional generative adversarial networks (GANs) but also perfectly restore the physical continuity of the signal.
[0029] Achieving adaptive denoising and feature regularization significantly improves the robustness of downstream models: During the dimensionality reduction compression and decoding reconstruction process of the masked autoencoder, the network tends to reconstruct structured effective vehicle vibration stripes, thus naturally filtering and smoothing out the inherent random high-frequency environmental noise of distributed fiber optic devices (DAS / DVS / OFDR). Injecting these denoised reconstructed samples with rich feature variations into the training set plays a strong role in data regularization. Experimental verification shows that the enhancement method described in this invention significantly improves the feature extraction capability, generalization capability, and robustness of downstream instance segmentation and object detection models when facing severe noise interference in real road networks and cross-sensor applications. Attached Figure Description
[0030] Figure 1 This is a flowchart of the fiber optic traffic signal data enhancement method based on a masked autoencoder as described in this invention;
[0031] Figure 2 This is a diagram illustrating the network architecture and pixel-level reconstruction principle of an asymmetric masked autoencoder-decoder.
[0032] Figure 3 This is a schematic diagram illustrating the visual effects of data augmentation using the method of the present invention in a specific embodiment;
[0033] Figure 4 This is a comparative schematic diagram of the feature heatmap (Grad-CAM) of the downstream sensing model in a specific embodiment when processing optical fiber signals. Detailed Implementation
[0034] 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.
[0035] Specific Implementation Method 1: Combination Figure 1 and Figure 2 As shown, this invention provides a fiber optic traffic signal data enhancement method based on a masked autoencoder, comprising:
[0036] The system acquires raw spatiotemporal vibration signals collected by distributed fiber optic sensing devices when a vehicle is traveling in a traffic network, generates a single-channel standardized two-dimensional spatiotemporal matrix, and then configures the original artificial fine-grained annotation set; the standardized two-dimensional spatiotemporal matrix is a waterfall plot grayscale image.
[0037] The standardized two-dimensional spatiotemporal matrix is divided into multiple regular and non-overlapping image patches to form a serialized input containing global spatiotemporal information. To simulate severe signal interference and local data loss in real fiber optic application scenarios, a high proportion of random masking is applied to the image patches to obtain visible block sequences and masked block sequences.
[0038] The visible block sequence and classification token ([CLS] Token) are input into a heavy encoder based on a self-attention mechanism for feature mapping to obtain a low-dimensional latent feature representation vector;
[0039] The low-dimensional latent feature representation vector and the masking markers corresponding to the masking block sequence are concatenated and combined according to their original spatial positions to obtain the concatenated complete sequence. Then, a lightweight decoder is used to optimize the concatenated complete sequence by minimizing the mean square error (Masked MSE Loss) of the region corresponding to the masking block sequence. Combined with the contextual semantic information of the visible block sequence, pixel-level signal reconstruction is performed to obtain the reconstructed spatiotemporal signal image.
[0040] The original set of fine-grained artificial annotations is mapped onto the reconstructed spatiotemporal signal image to obtain twin-enhanced samples with high-precision labels.
[0041] In this embodiment, based on a preset mask ratio, all image blocks are divided into visible block sequences and masked block sequences; wherein, the visible block sequences retain only a small number of sparse fiber vibration signal segments.
[0042] Furthermore, combined with Figure 2 As shown, the heavyweight encoder captures the inherent global topology and long-range spatiotemporal dependence of the original spatiotemporal vibration signal based on the visible block sequence, and obtains a low-dimensional latent feature representation vector.
[0043] The encoder only processes the visible signal segments that are not masked, forcing the deep learning network to bypass local pixel textures in order to capture the inherent global topological structure and long-range spatiotemporal dependencies of the fiber optic vibration signal, and finally outputs a low-dimensional latent representation vector.
[0044] The lightweight decoder infers and completes the missing fiber optic signal regions corresponding to the masked block sequence based on the contextual semantic information of the visible block sequence, thereby obtaining the reconstructed spatiotemporal signal image. The reconstructed spatiotemporal signal image not only restores the physical topological continuity of the vehicle vibration main ridge line, but also filters out random high-frequency background noise from the equipment in the original image.
[0045] Based on the principle of semantic inheritance, since the reconstructed image strictly maintains the spatial semantic layout and geometric shape of the vehicle trajectory in the original image, the system directly maps (reuses) the original fine-grained annotation set onto the reconstructed image without introducing any additional manual annotation costs, forming twin augmented sample pairs with high-precision labels.
[0046] The original spatiotemporal vibration signal is aggregated with the twin-enhanced sample to construct an augmented training set that is multiplied. :
[0047] ,
[0048] In the formula, N represents the number of original spatiotemporal vibration signals. For the i-th original spatiotemporal vibration signal, Let i be the set of original artificial fine-grained annotations corresponding to the i-th original spatiotemporal vibration signal. For the i-th reconstructed spatiotemporal signal image, The reconstruction mapping function is the masked autoencoder after convergence trained on large-scale unsupervised data. The masked autoencoder is a complete asymmetric network consisting of a heavyweight encoder and a lightweight decoder. Let be the random binary masking matrix corresponding to the random masking rate.
[0049] As an example, the random masking rate is not less than 75%.
[0050] In this embodiment, the enhanced training set Used for training downstream multi-task deep learning perception models.
[0051] As an example, the multi-task deep learning perception model includes an instance segmentation model, an object detection model, and a semantic segmentation network model.
[0052] As an example, the distributed fiber optic sensing device acquires data including DAS, DVS, and OFDR.
[0053] In this embodiment, the original spatiotemporal vibration signal is preprocessed by denoising and downsampling to generate a single-channel standardized two-dimensional spatiotemporal matrix.
[0054] The original manual fine-grained annotation set described in this embodiment includes instance segmentation mask and bounding box (Bbox).
[0055] This invention also provides a distributed fiber optic traffic signal data enhancement system based on a masking autoencoder and self-supervised reconstruction, comprising: a preprocessing and segmentation module for acquiring a spatiotemporal waterfall plot of fiber optic sensing and its manual annotations, and dividing it into regular, non-overlapping image blocks; a random masking module for performing high-proportion random masking on the image blocks, simulating signal loss and separating visible blocks and masked block sequences; a feature encoding module with a built-in heavyweight self-attention encoder, which processes only the visible blocks to extract latent feature representations with global topological priors; a reconstruction and decoding module with a built-in lightweight decoder for concatenating the latent feature representations and masking labels and performing pixel-level reconstruction, outputting a topologically continuous and smoothly denoised reconstructed signal image; and a label lossless reuse and training module for directly mapping the original labels to the reconstructed image to form twin augmented samples, and combining them to construct an augmented training set to train the downstream traffic perception network. Specific Implementation
[0056] The following, in conjunction with the accompanying drawings, provides a method and system for enhancing distributed fiber optic traffic signal data based on masked autoencoders (MAE) and self-supervised reconstruction. This method is applied to distributed fiber optic traffic sensing edge computing nodes or cloud servers. This embodiment is consistent with existing conventional application scenarios, using a 1km long two-way highway as the specific monitoring scenario, where distributed vibration sensing (DVS) devices are deployed on the roadside to perceive traffic flow status.
[0057] Unlike traditional physical augmentation methods based on pixel cropping and stitching, this embodiment aims to leverage MAE's powerful contextual reasoning and structural repair capabilities to generate noise-resistant and robust Siamese augmented samples without altering the original spatiotemporal signal semantic space layout. The specific execution flow of this augmentation method includes the following steps:
[0058] Step 1: Acquire labeled fiber optic spatiotemporal signals and divide them into image blocks. For example, at time T, a segment of raw fiber optic vibration signal containing vehicle driving status was acquired using a roadside DVS device.
[0059] Step 1.1: Obtain the original DVS signal matrix. In this embodiment, the spatial resolution of the DVS device is set to 1 meter, and the monitoring distance is 1 km (corresponding to 1000 spatial sampling points); the vibration signal within 10 seconds is extracted on the time axis (corresponding to 5000 time sampling points). To adapt to the visual Transformer (ViT) input dimension of the Masked Autoencoder (MAE) standard and to strictly distinguish it from the downsampling strategy of heterogeneous devices, bilinear interpolation and adaptive grid pooling are performed along the time and spatial axes to accurately compress and transform the original long-range high-frequency signal into a matrix of size 10 ... Normalized two-dimensional spatiotemporal matrix of pixels (i.e., a single-channel waterfall grayscale image). Simultaneously, extract this matrix. The corresponding set of original manual fine-grained annotations Assume that it contains the bounding box coordinates (Bbox) and pixel-level mask of a vehicle traveling in the forward direction.
[0060] Step 1.2: As Figure 2 As shown, the size is Two-dimensional spacetime matrix The image is divided into multiple regular, non-overlapping image patches. In this embodiment, the image patch size is set to [size missing]. Pixels, thus flattening and segmenting the entire spatiotemporal image. A sequence of image patches.
[0061] Step 2: High-proportion random masking operation and sequence separation. To simulate severe signal breakage and data loss caused by poor road surface coupling or strong transient environmental noise (such as strong winds or construction) in real fiber optic application scenarios, a high-proportion random masking strategy is applied to the above 196 image blocks. Combined with... Figure 3 The original input image shown in (a) is used to generate an image in this embodiment. Figure 3 The binarized random mask shown in (b) has a mask ratio set to an extremely high 75%. Based on the uniform distribution of this mask, the system randomly discards 147 image patches (i.e., the masked patch sequence), retaining only the remaining 49 image patches (i.e., the visible patch sequence). At this point, as... Figure 3 As shown in (c), the highly fragmented masking effect after applying the mask results in an extremely sparse and fragmented input image. Conventional local convolutional networks are unable to identify the continuous vehicle vibration features within it. Finally, we obtain... Figure 3 The output image shown in (d) is the result of decoding and reconstruction by an autoencoder.
[0062] Step 3: Latent Feature Extraction Based on Asymmetric Heavyweight Encoder. The 49 visible patch sequences separated in Step 2 are linearly projected and coupled with positional embeddings that preserve the original spatiotemporal location information. A classification token ([CLS] Token) is then added, and all are uniformly input into a heavyweight encoder based on a multi-head self-attention mechanism. Since the encoder only processes 25% of the visible signal segments, the computational complexity is greatly reduced. More importantly, this information bottleneck forces the deep learning network to overcome the superficial local high-frequency noise and deeply mine and extract the inherent global topological structure and long-range spatiotemporal dependencies of fiber optic vehicle vibration signals (such as V-shaped or parabolic ripples), ultimately outputting a high-dimensional latent feature representation vector rich in topological priors.
[0063] Step 4: Pixel-level reconstruction and topology repair based on a lightweight decoder.
[0064] Step 4.1: As Figure 2 and Figure 3 As shown, the latent feature representation vector output in step 3 is re-assembled and aligned with the 147 mask tokens (i.e., a set of shared, learnable vectors) discarded in step 2 according to the spatial location index of the original image patch, to restore a complete sequence of length 196.
[0065] Step 4.2: Add the complete decoder position encoding to the concatenated complete sequence and input it into the Light Decoder.
[0066] Step 4.3: Based on the contextual semantics and global topological features of the 25% visible portion, the decoder infers and accurately predicts the pixel values of the masked 75% region. The final output size is also [size missing]. Reconstructed spatiotemporal signal image Its visual effects are like Figure 3 As shown in (d). Compare. Figure 3 (a) and Figure 3 As can be clearly seen in (d), the reconstructed image It not only perfectly repaired the vehicle vibration main ridge line that might have broken due to environmental interference, restoring physical topological continuity; at the same time, due to the information compression characteristics of the autoencoder, it naturally filtered out the original image. The high-frequency background noise of the random DVS device is used to achieve adaptive denoising and smoothing regularization effects.
[0067] Step 5: Annotation lossless reuse and hybrid augmentation dataset construction.
[0068] Step 5.1: Zero-cost annotation inheritance. Since the above asymmetric MAE reconstruction process strictly follows pixel-level restoration based on position encoding, the reconstructed image... Spatial semantic layout, geometric scale and original image of vehicle trajectory To maintain absolute consistency, the system directly extracts the original fine-grained annotation set from step 1.1. (that is, originally in) Bbox and Mask in coordinate system can be directly mapped and reused to the reconstructed image without any coordinate transformation or manual verification. A pair of high-quality twin augmented samples was seamlessly generated. .
[0069] Step 5.2: Traverse the real distributed fiber optic traffic dataset, continuously generate corresponding self-supervised reconstruction twin samples using the above steps, and mix them with the original dataset to construct a large-scale augmented training set that is multiplied. .
[0070] Step 5.3: Input the enhanced training set into a downstream multi-task deep learning perceptual model (such as a fiber optic vehicle instance segmentation network based on Mask R-CNN) for training. Because the training set incorporates a large number of reconstructed samples with altered noise distribution characteristics after topological repair, the downstream model is forced to learn more robust and essential vehicle vibration physical characteristics. Engineering verification shows that after enhancement using this method, the model achieves a breakthrough improvement in both detection recall and segmentation accuracy when dealing with low signal-to-noise ratio fiber optic signals under real-world complex road conditions.
[0071] To intuitively verify the effectiveness and mechanism of the enhancement method of the present invention, combined with Figure 4 The model feature heatmap (Grad-CAM) shown is used for comparison and explanation: For example Figure 4 Image (a) shows the original spatiotemporal signal image of the input fiber optic signal, which includes a tilted effective vibration ridge of the vehicle and a large amount of vertical high-frequency environmental noise. Figure 4 As shown in (b), the feature heatmap of the baseline model trained using only limited raw real data is highly susceptible to interference from complex environments. The model's attention becomes severely scattered, with numerous highly activated regions—i.e., the highlighted / warm-colored parts in the image—mistakenly falling onto irrelevant background noise textures, making it difficult for the model to establish robust classification and segmentation boundaries. Conversely, as... Figure 4As shown in (c), this is a heatmap of model features after training with the enhanced training set containing a large number of topology repair and adaptive denoising twin samples, as described in this invention. The internal decision-making mechanism of the downstream perception model has been fundamentally optimized. Its highly activated feature regions are highly concentrated and fit perfectly with the main ridge line of the actual vehicle vibration. This visualization result fully demonstrates that the enhanced data generated by this invention forces the deep learning network to escape the trap of local noise and truly grasp the global topology and physical continuity of the fiber optic traffic signal, thus exhibiting extremely strong noise robustness under poor signal-to-noise ratios.
[0072] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.
Claims
1. A method for enhancing fiber optic traffic signal data based on a masked autoencoder, characterized in that... include, The system acquires raw spatiotemporal vibration signals collected by distributed fiber optic sensing devices during vehicle operation, generates a standardized two-dimensional spatiotemporal matrix for a single channel, and then configures the original set of fine-grained artificial annotations. The standardized two-dimensional spatiotemporal matrix is divided into multiple regular and non-overlapping image blocks, and a high proportion of random masking is applied to obtain visible block sequences and masked block sequences. The visible block sequence and classification label are input into a heavyweight encoder based on a self-attention mechanism for feature mapping to obtain a low-dimensional latent feature representation vector; The low-dimensional latent feature representation vector and the masking markers corresponding to the masking block sequence are concatenated and combined according to their original spatial positions to obtain the concatenated complete sequence. Then, a lightweight decoder is used to optimize the concatenated complete sequence by minimizing the mean square error of the region corresponding to the masking block sequence. Combined with the contextual semantic information of the visible block sequence, pixel-level signal reconstruction is performed to obtain the reconstructed spatiotemporal signal image. The original set of fine-grained artificial annotations is mapped onto the reconstructed spatiotemporal signal image to obtain twin-enhanced samples with high-precision labels.
2. The fiber optic traffic signal data enhancement method based on a masked autoencoder according to claim 1, characterized in that, The heavyweight encoder captures the inherent global topology and long-range spatiotemporal dependencies of the original spatiotemporal vibration signal based on visible block sequences, and obtains a low-dimensional latent feature representation vector.
3. The fiber optic traffic signal data enhancement method based on a masked autoencoder according to claim 2, characterized in that, The lightweight decoder infers and completes the corresponding region of the masked block sequence based on the contextual semantic information of the visible block sequence, thereby obtaining the reconstructed spatiotemporal signal image.
4. The fiber optic traffic signal data enhancement method based on a masked autoencoder according to claim 1, characterized in that, The original spatiotemporal vibration signal is aggregated with the twin-enhanced sample to construct an enhanced training set. : ， In the formula, N represents the number of original spatiotemporal vibration signals. For the i-th original spatiotemporal vibration signal, Let i be the set of original artificial fine-grained annotations corresponding to the i-th original spatiotemporal vibration signal. For the i-th reconstructed spatiotemporal signal image, The reconstruction mapping function is the masked autoencoder after convergence trained on large-scale unsupervised data. The masked autoencoder is a complete asymmetric network consisting of a heavyweight encoder and a lightweight decoder. Let be the random binary masking matrix corresponding to the random masking rate.
5. The fiber optic traffic signal data enhancement method based on a masked autoencoder according to claim 4, characterized in that, The random masking rate is not less than 75%.
6. The fiber optic traffic signal data enhancement method based on a masked autoencoder according to claim 4, characterized in that, The enhanced training set Used for training downstream multi-task deep learning perception models.
7. The fiber optic traffic signal data enhancement method based on a masked autoencoder according to claim 6, characterized in that, The multi-task deep learning perception model includes an instance segmentation model, an object detection model, and a semantic segmentation network model.
8. The fiber optic traffic signal data enhancement method based on a masked autoencoder according to claim 1, characterized in that, The distributed optical fiber sensing device acquires data including DAS, DVS, and OFDR.
9. The fiber optic traffic signal data enhancement method based on a masked autoencoder according to claim 1, characterized in that, The original spatiotemporal vibration signal is preprocessed by denoising and downsampling to generate a single-channel standardized two-dimensional spatiotemporal matrix.
10. The fiber optic traffic signal data enhancement method based on a masked autoencoder according to claim 1, characterized in that, The original set of fine-grained manual annotations includes instance segmentation masks and bounding boxes (Bboxes).