A physical a priori distributed fiber sensing real-time device and method

Abstract

This invention relates to the fields of fiber optic sensing technology and deep learning technology, and particularly to a physical prior distributed fiber optic sensing real-time device and method. The method includes capturing optical signals transmitted in long-distance optical fibers, converting the optical signals into digital time-series signals, and constructing a spatiotemporal matrix from the digital time-series signals; constructing a sensing model, and guiding the training of the sensing model by extracting frequency domain prior knowledge from optical signal processing, enabling the sensing model to extract accurate localization frequency domain information; and running the trained sensing model on an accelerator to process the spatiotemporal matrix to identify vibration events along the long-distance optical fiber externally online. This invention can perform online analysis of optical signals in long-distance monitoring scenarios to reduce latency.

Description

Technical Field

[0001] This invention relates to the fields of fiber optic sensing technology and deep learning technology, and in particular to a physical prior distributed fiber optic sensing real-time device and method. Background Technology

[0002] Distributed fiber optic sensing technology is a key means of measuring and monitoring physical information along fiber optic cables and has been widely used in the monitoring and maintenance of long-distance infrastructure. By capturing scattered light signals and analyzing them using models, distributed fiber optic sensing technology can perceive changes in physical parameters around the fiber in real time. However, with the extension of the monitoring span and the increase in fiber length, the amount of raw data generated by the sensing system has exploded, posing a severe challenge to the processing architecture. On the one hand, the expanded monitoring range means that the environment in which the fiber is located is more diverse, and traditional signal processing algorithms struggle to accurately extract features from raw signals with low signal-to-noise ratios and complex interference. On the other hand, although the high density of sensor data makes it possible to introduce cutting-edge algorithms such as deep learning, the lack of strategies that can simultaneously handle massive data throughput and complex signal processing quality means that most current optical signal analysis is performed offline, which makes true real-time monitoring difficult.

[0003] Offline analysis refers to collecting optical signal data over a period of time (such as a week) and then processing the large amount of optical signal data in a centralized manner to sense changes in physical parameters around the optical fiber at a certain moment in the previous period (a week). Even if parameter changes are detected, this analysis only reflects the past state and has a significant lag. This characteristic makes offline analysis unable to meet the application scenarios with high timeliness, such as geological disaster early warning and real-time monitoring of safety accidents. Furthermore, offline analysis generally relies on deep learning models and mainstream computing platforms. While deep learning models possess powerful feature extraction capabilities, their deployment is highly dependent on high-performance computing devices, and the algorithms available in actual engineering settings are often very limited. Existing mainstream computing platforms, such as central processing units (CPUs) or graphics processing units (GPUs), are often limited by serial instruction execution modes or general architectural designs when handling such tasks. This results in their parallel processing performance not being deeply optimized for the specific needs of distributed sensing. This hardware-level limitation makes it difficult for the system to maintain low-latency operation under massive amounts of data. In long-distance monitoring scenarios, the technical bottleneck of the mismatch between the high-concurrency, massive data throughput requirements generated by the sensing system and the limited processing capabilities of the underlying computing hardware greatly restricts the widespread adoption of distributed fiber optic sensing technology in more diverse scenarios.

[0004] In summary, offline analysis has limited applicability in long-distance monitoring scenarios. How to achieve efficient online analysis to reduce latency and promote the widespread adoption of distributed fiber optic sensing technology in a wider range of scenarios (especially in high-time-sensitivity applications) is a problem that needs to be solved.

[0005] Therefore, overcoming the shortcomings of the existing technology is an urgent problem to be solved in this technical field. Summary of the Invention

[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention proposes a physical prior distributed optical fiber sensing real-time device and method, which can realize online analysis of optical signals in long-distance monitoring scenarios to reduce latency.

[0007] The embodiments of the present invention adopt the following technical solutions: In a first aspect, the present invention provides a physical prior distributed optical fiber sensing real-time method, specifically: capturing optical signals transmitted in long-distance optical fibers, converting the optical signals into digital time-series signals, and constructing the digital time-series signals into a spatiotemporal matrix; A perception model is constructed, and the training of the perception model is guided by extracting frequency domain prior knowledge from optical signal processing, so that the perception model has the ability to extract accurate localization frequency domain information. The accelerator runs the trained perception model and processes the spatiotemporal matrix to identify vibration events along the long-distance optical fiber in the online manner.

[0008] Preferably, the accelerator includes an FPGA, and the method further includes: The multiplication and accumulation operations in the perception model are replaced with logical shift and addition operations, and hardware acceleration is achieved through a pipelined parallel structure on the FPGA.

[0009] Preferably, the method further includes: The perception model includes a first convolutional block and a second convolutional block connected in series, and a recognizer connected to the output of the second convolutional block. The first convolutional block includes at least one depth-separable convolutional layer, and the second convolutional block includes at least one depth-separable convolutional layer, a max pooling layer, and an average pooling layer.

[0010] Preferably, the method for training the perception model includes: Construct a training set, which includes a spatiotemporal matrix generated by optical signal processing and its corresponding real vibration events; The spatiotemporal matrix in the training set is input into the perception model to obtain the prediction result output by the recognizer and the intermediate features extracted by the first convolutional block. Perform a discrete Fourier transform on the spatiotemporal matrix in the training set along the time axis to obtain the physical frequency domain features corresponding to the spatiotemporal matrix; Based on the prediction results, actual vibration events, intermediate features, and physical frequency domain features, the total loss value is calculated according to the defined total loss function. Based on the total loss value, the network parameters are updated using the backpropagation algorithm. The network parameters are updated repeatedly based on the total loss value calculated from the spatiotemporal matrix in the training set until the preset convergence condition is met, and the trained perception model is obtained.

[0011] Preferably, the total loss function includes matching loss and perceptual loss, wherein the matching loss is used to align the kernel weights of the first convolutional block with the structure of the frequency basis function, and forces the first convolutional block to simulate the frequency selection characteristics of Fourier analysis during gradient descent.

[0012] Preferably, the total loss function is defined as follows: ; in, Represents the total loss value. Represents the matching loss value. Represents the perceived loss value. This represents the domain weight coefficient.

[0013] Preferably, the matching loss is defined as follows: ; in, Represents the matching loss value. This represents the intermediate features extracted from the first convolutional block. It represents the physical frequency domain characteristics.

[0014] Preferably, the step of inputting the spatiotemporal matrix from the training set into the perceptual model to obtain the prediction result output by the recognizer and the intermediate features extracted by the first convolutional block includes: The first convolutional block is used to perform preliminary feature extraction on the spatiotemporal matrix to obtain intermediate features with frequency domain information; The intermediate features are extracted using a second convolutional block to obtain a feature map. The feature map is then flattened and input into the recognizer to obtain a prediction result.

[0015] In a second aspect, the present invention provides a physical prior distributed fiber optic sensing real-time device, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the processor for performing the physical prior distributed fiber optic sensing real-time method of the first aspect.

[0016] Thirdly, the present invention also provides a non-volatile computer storage medium storing computer-executable instructions that are executed by one or more processors to perform the physical prior distributed fiber optic sensing real-time method described in the first aspect.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: it constructs a lightweight sensing model and improves the feature extraction capability and generalization performance of the sensing model in complex environments based on a uniquely designed training method, thereby realizing the application of the lightweight model in fiber optic sensing. It also breaks through the performance bottleneck of traditional general-purpose processors through embedded hardware optimization. Based on the lightweight sensing model and embedded hardware, it performs online analysis of optical signals. This invention significantly improves the system's processing efficiency and reduces response latency, solving the practical application problem of distributed fiber optic sensing technology in large-scale monitoring and long-distance applications. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This invention provides a distributed optical fiber sensor signal real-time processing system. Figure 2 This is a flowchart illustrating a physical prior distributed optical fiber sensing real-time method provided in an embodiment of the present invention. Figure 3 This is a flowchart illustrating a method for training a perception model according to an embodiment of the present invention; Figure 4 This is a flowchart of a method for training a perception model provided in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the specific implementation method of step 202 provided in an embodiment of the present invention; Figure 6 This is a structural diagram of a physical prior distributed optical fiber sensing real-time method provided in an embodiment of the present invention; Figure 7 A schematic diagram of the structure of a physical prior distributed fiber optic sensing real-time device provided in an embodiment of the present invention; The reference numerals in the attached figures are as follows: 21: Processor; 22: Memory. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] Unless the context otherwise requires, throughout the specification and claims, the term "comprising" is interpreted as openly inclusive, meaning "including, but not limited to." In the description of the specification, terms such as "one embodiment," "some embodiments," "exemplary embodiment," "example," "specific example," or "some examples" are intended to indicate that a particular feature, structure, material, or characteristic associated with that embodiment or example is included in at least one embodiment or example of this disclosure. The illustrative representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics mentioned may be included in any suitable manner in any one or more embodiments or examples; that is, although they may be incorporated into embodiments or examples using the above terms for reasons such as order and position, it does not limit them to be incorporated in combination by a single embodiment or example.

[0022] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, for example, the description may use the prefix "A" or "B" to describe the same type of nouns as two independent entities. In this case, the corresponding features defined with "A" and "B" are used only to distinguish between similar entities and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features.

[0023] In the description of this invention, the expression “A and / or B” (where A and B are used to formally represent specific features) will be used. The corresponding expression includes the following three combinations: only A, only B, and a combination of A and B.

[0024] As used in this invention, “about,” “approximately,” or “approximately” includes the stated value and the average value within an acceptable range of deviation from a particular value, wherein the acceptable range of deviation is determined by a person skilled in the art taking into account the measurement under discussion and the error associated with the measurement of the particular quantity (i.e., the limitations of the measurement system).

[0025] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0026] Example 1: This invention achieves online analysis of optical signals by integrating a lightweight sensing model and a hardware accelerator. Online analysis allows users to preset a data acquisition window; when the accumulated optical signal samples reach a preset threshold, the algorithm model automatically activates, providing real-time sensing of changes in physical parameters along the optical fiber over a preceding time period (e.g., minutes or ten minutes). Compared to offline analysis, this online mechanism eliminates data waiting time, significantly reduces system latency, and greatly improves response speed, meeting the stringent real-time monitoring requirements of time-sensitive scenarios. Furthermore, because online analysis employs a preset trigger threshold mechanism, the data volume processed in a single operation is far smaller than that of offline analysis. This mode effectively avoids the storage and computational pressure caused by explosive growth in raw data, significantly improving the system's load capacity.

[0027] This invention also provides a distributed optical fiber sensor signal real-time processing system for use in conjunction with the distributed optical fiber sensor signal real-time processing method, see below. Figure 1 A distributed optical fiber sensing signal real-time processing system includes a physical sensing layer, an edge computing layer, and an application terminal layer. The physical sensing layer includes optical sensing links (such as lasers or modulators), sensing optical fibers, and photodetectors. The edge computing layer includes high-speed analog-to-digital conversion and signal processing units. The application terminal layer includes a monitoring host computer.

[0028] The physical sensing layer utilizes a distributed fiber optic sensing system to capture optical signals transmitted over long distances in optical fibers. Specifically, in the optical sensing link, continuous light waves are modulated into extremely short probe light pulses, which are then injected into the sensing fiber. As the light pulses propagate within the fiber, they interact with the microstructure of the fiber material, generating weak scattered light. This scattered light carries environmental information such as temperature, strain, or vibration distributed along the fiber and returns along the original path or a specific path. The returned weak scattered light signals enter a highly sensitive photodetector and are converted into corresponding current signals. After amplification and conditioning, the current signals are converted into analog electrical signals, which are then transmitted to the edge computing layer. The edge computing layer receives the analog electrical signals converted from the photodetector as input and processes them through a high-speed analog-to-digital converter. The converter (ADC) converts the signal into a digital time-series signal. The signal processing unit deployed on the accelerator receives the digital time-series signal from the ADC, loads the trained perception model, processes the digital time-series signal to accurately identify vibration events along the long-distance optical fiber, and transmits the vibration events to the application terminal layer. The application terminal layer is responsible for the system's human-computer interaction, data communication, and result presentation. After the monitoring host computer receives the vibration event, it provides a real-time visual display and alarm for the user's decision-making.

[0029] like Figure 2 As shown, Embodiment 1 of the present invention provides a physical prior distributed fiber optic sensing real-time method, which specifically includes the following steps: Step 101: Capture the optical signal transmitted in the long-distance optical fiber, convert the optical signal into a digital time-series signal, and construct the digital time-series signal into a spatiotemporal matrix.

[0030] This step is implemented by the physical sensing layer and the edge computing layer in the distributed optical fiber sensing signal real-time processing system. The physical sensing layer uses the distributed optical fiber sensing system to capture optical signals transmitted in long-distance optical fibers; the edge computing layer converts the optical signals into digital timing signals through an ADC.

[0031] The method for constructing the digital time-series signal into a spatiotemporal matrix is ​​as follows: the one-dimensional long sequence of digital time-series signal is sliced ​​according to a preset spatial resolution and time window, thereby constructing the spatiotemporal matrix. Where T represents the number of time steps, S represents the number of spatial sampling points, and the spatiotemporal matrix is... Each row in the array represents the instantaneous state distribution of all spatial sampling points at a specific moment, and each column represents the complete signal sequence of a specific spatial location evolving over time. For example, the acquired one-dimensional raw signal can be sliced ​​into... The spatiotemporal matrix input (representing the vibration characteristics of a 12.5-meter fiber within 256 milliseconds).

[0032] Step 102: Construct a perception model. Use prior knowledge in the frequency domain of optical signal processing to guide the training of the perception model, so that the perception model has the ability to extract accurate localization frequency domain information.

[0033] To achieve the goal of real-time processing of sensor signals, this invention employs a depth-optimized lightweight model at the processing algorithm level. The constructed perception model includes a first convolutional block and a second convolutional block connected in series, as well as a recognizer connected to the output of the second convolutional block. The first convolutional block includes at least one depth-separable convolutional layer, and the second convolutional block includes at least one depth-separable convolutional layer, a max pooling layer, and an average pooling layer.

[0034] Compared to standard convolution, this invention uses depthwise separable convolution as the core operator. This operator decomposes the convolution process into depthwise convolution and pointwise convolution. Depthwise convolution independently applies a single filter to each channel of the input to extract spatial features; pointwise convolution uses... The convolutional kernel linearly combines features from different channels; this design reduces computational complexity from... Significantly reduced to This significantly reduces the consumption of multipliers and memory resources on the FPGA, achieving higher parallelism, throughput, and lower power consumption at the hardware level, where K is the convolution kernel size. Input the number of channels. This represents the number of output channels.

[0035] However, lightweight models generally suffer from limited processing capacity and low generalization, making it difficult to achieve a balance between accuracy and speed in long-distance monitoring applications.

[0036] To address the issue of low generalization ability in lightweight models, this invention proposes a cross-domain distillation method that does not require a teacher network. This method guides the training of a time-domain convolutional neural network by extracting frequency domain prior knowledge from physical signal processing. This enables the model to learn features with clear physical meaning and strong generalization ability while maintaining a very low parameter scale. In other words, this invention is based on the idea of ​​a teacher model guiding a student model, enabling a perception model capable of performing recognition tasks (recognizing vibration events along the exterior of long-distance optical fibers) to extract frequency domain information. However, during training, this invention does not require the use of a separate teacher model to guide the training of the student model. Instead, it constructs two feature extraction paths in parallel and adaptively designs matching loss functions for intermediate features and frequency domain features. By combining the designed matching loss function with the perception loss function related to the recognition task, the perception model can extract features containing frequency domain information on the one hand, and achieve good recognition accuracy for the recognition task on the other.

[0037] The finally trained perception model is deployed on an accelerator for acceleration.

[0038] Step 103: The accelerator runs the trained perception model and processes the spatiotemporal matrix to identify vibration events along the long-distance optical fiber in the online manner.

[0039] The accelerator is either an FPGA or a GPU. In this embodiment, the following description of the accelerator will use an FPGA as an example.

[0040] One advantage of FPGA-based hardware architecture is its designability. Using FPGA as a hardware platform can provide hardware acceleration support for the rapid operation of perception models.

[0041] There are two main reasons why FPGAs can accelerate the rapid operation of sensing models. First, the hardware architecture of FPGAs can be optimized and adapted to specific algorithm models. Customizing the FPGA hardware architecture based on the sensing model's structure can significantly improve the model's operating efficiency. Compared to deploying sensing models on CPUs or GPUs, using FPGAs for acceleration achieves faster inference speeds and higher energy efficiency. Second, FPGAs have low power consumption requirements. Complete servers or general-purpose computer systems rely on robust power grid infrastructure, while FPGAs can operate without this infrastructure, significantly reducing power consumption and greatly expanding their deployment scenarios. For example, in distributed fiber optic sensing applications, in harsh environments lacking stable power support, such as geological monitoring or mine inspection, FPGAs only require 12V 5A power (total power consumption of approximately 60W). This not only demonstrates strong adaptability but can even be driven by portable power banks, truly enabling flexible edge deployment.

[0042] In this embodiment, a lightweight sensing model is constructed, and based on a uniquely designed training method, the feature extraction capability and generalization performance of the sensing model in complex environments are improved, thereby realizing the application of the lightweight model in fiber optic sensing. Furthermore, the performance bottleneck of traditional general-purpose processors is overcome through embedded hardware optimization. Based on the lightweight sensing model and embedded hardware, online analysis of optical signals is performed. This invention significantly improves the system's processing efficiency and reduces response latency, solving the practical application problem of distributed fiber optic sensing technology in large-scale monitoring and long-distance applications.

[0043] Furthermore, to achieve low-latency processing at resource-constrained edge computing environments, this invention employs an FPGA-optimized quantization and parallel processing strategy. Specifically, it replaces the multiplication-accumulation operations in the perceptual model with logic shift and addition operations, and accelerates this through a pipelined parallel structure on the FPGA. This invention uses a special quantization scheme to replace traditional multiplication operations in neural networks with logic shift and addition operations. This transformation allows the perceptual model to utilize the basic logic units of the FPGA, including look-up tables (LUTs) and flip-flops (FFs), without consuming the limited resources of digital signal processing (DSP) chips. Additionally, by eliminating the DSP, the perceptual model achieves higher parallelism, significantly reducing inference latency. The three-layer depthwise separable convolutional structure of the perceptual model allows each layer to build an independent hardware-accelerated pipeline, thereby enabling point-by-point real-time processing of long sequences of data.

[0044] Based on the perceptual model constructed in step 102, a possible implementation is as follows: A three-layer depthwise separable convolutional layer is used as the core, with the number of channels increasing layer by layer from 8 to 32 to preserve data dimensionality. The first layer forms the first convolutional block, and the second and third layers form the second convolutional block, in conjunction with a fully connected layer as the recognizer. The entire model has a total of 4141 parameters, making it a highly lightweight model that significantly reduces computational cost while maintaining high feature extraction capabilities.

[0045] like Figure 3 As shown, this embodiment of the invention provides a method for training a perceptual model, which specifically includes the following steps: Step 201: Construct a training set, which includes a spatiotemporal matrix generated by optical signal processing and its corresponding real vibration events.

[0046] A pre-defined length (e.g., 50 meters) of fiber optic cable is buried in the ground. Typical vibration events are simulated around this section. Two sets of data are collected at different times and locations. One set of data is used as unknown data for the sensing model to test its out-of-distribution generalization, i.e., the performance of the sensing model in an unknown environment. The other set of data is used as data for model training, validation, and in-distribution generalization testing.

[0047] The training set includes a spatiotemporal matrix generated by optical signal processing and its corresponding real vibration event labels. The spatiotemporal matrix is ​​obtained by processing the collected optical signal based on the method described in step 101. The real vibration events are used to reflect physical changes that occur outside the optical fiber, such as human excavation, mechanical construction, vehicle traffic, pedestrian trampling, or natural geological disasters.

[0048] Step 202: Input the spatiotemporal matrix from the training set into the perception model to obtain the prediction results output by the recognizer and the intermediate features extracted by the first convolutional block.

[0049] Steps 202 and 203 are two parallel paths, namely the time-domain path and the frequency-domain path. The content obtained from the time-domain path is used as the student content, and the content obtained from the frequency-domain path is used as the reference content. This invention does not require the cross-domain distillation method of the teacher network. Instead, it guides the training of the time-domain convolutional neural network by extracting frequency-domain prior knowledge from physical signal processing. That is, the reference content is used to guide the student content, so that the perceptual model can ultimately learn features with clear physical meaning and strong generalization ability.

[0050] The temporal path is essentially the process by which the perceptual model processes the spatiotemporal matrix. The resulting prediction is the final output of the perceptual model, and the intermediate matrix is ​​obtained by extracting local features of the spatiotemporal matrix from the first convolutional block.

[0051] Step 203: Perform a discrete Fourier transform on the spatiotemporal matrix in the training set along the time axis to obtain the physical frequency domain features corresponding to the spatiotemporal matrix.

[0052] The frequency domain path is the process of extracting physical frequency domain features through Fourier transform, specifically: for the same spatiotemporal matrix (The spatiotemporal matrix in step 202) is subjected to a Discrete Fourier Transform along the time axis to obtain the physical frequency domain features corresponding to the spatiotemporal matrix. This approach utilizes prior physical knowledge as a reference, eliminating the need for additional training of a teacher model, thus enabling the distillation of teacherless networks.

[0053] Step 204: Based on the prediction results, actual vibration events, intermediate features, and physical frequency domain features, calculate the total loss value according to the defined total loss function.

[0054] The total loss function includes matching loss and perceptual loss, where the perceptual loss is task-related. (For example, cross-entropy loss). When calculating the total loss value, on the one hand, the perceptual loss is used to calculate the loss value between the predicted result and the actual vibration event, which is used to evaluate the gap between the final output capability of the perceptual model and the actual content; on the other hand, constrained optimization methods are used to extract intermediate features from temporal convolution. Physical frequency domain characteristics Alignment is performed to calculate the matching loss. The matching loss and the perceptual loss are then weighted and fused to construct the total loss function, which is defined as follows: ; in, Represents the total loss value. Represents the matching loss value. Represents the perceived loss value. Represents the domain weight coefficient. It can be determined using a grid search method.

[0055] The matching loss is used to align the kernel weights of the first convolutional block with the structure of the frequency basis function, and forces the first convolutional block to simulate the frequency selection characteristics of Fourier analysis during gradient descent, thereby capturing more robust vibrational features. The matching loss is defined as follows: ; in, Represents the matching loss value. This represents the intermediate features extracted from the first convolutional block. It represents the physical frequency domain characteristics.

[0056] Step 205: Based on the total loss value, update the network parameters through the backpropagation algorithm. Repeat the process of updating the network parameters based on the total loss value calculated from the spatiotemporal matrix in the training set until the preset convergence condition is met, and obtain the trained perception model.

[0057] The preset convergence condition is that the total loss value no longer decreases significantly or reaches the preset learning rate in multiple consecutive training rounds.

[0058] The training process is as follows: The Adam optimizer is used, with the initial learning rate set to 0.01. In the total loss function, the domain weight coefficients are determined through grid search. The value is 0.3. The model was trained using the above method, and the final model achieved a high accuracy of 99.49% in the in-distribution test. After being guided by frequency domain physics knowledge, the model maintained its in-distribution test performance, and its out-of-distribution generalization test results, that is, the recognition accuracy of the model in unknown environments, significantly improved from 51.93% to 95.72%.

[0059] After training, the perceptual model establishes a clear equivalent mapping mechanism between the single depth separable convolutional layer of the first convolutional block and the classical discrete Fourier transform physical operator. The equivalent mapping mechanism includes the structural alignment of the convolutional kernel weights and the frequency basis function, and the logical equivalence of feature decomposition and synthesis.

[0060] The method for aligning the structure of the convolution kernel weights with the frequency basis function is as follows: by constraining the cross-domain distillation loss function, the filter weights of this layer are forced to re-evolve into a form similar to the discrete Fourier transform basis function during the training process. This mapping mechanism makes the first convolution block no longer a black box feature extractor, but has the ability to perform spectral decomposition on the original time-domain optical signal, thus achieving consistency with the physical operator in mathematical logic.

[0061] The logical equivalence method of feature decomposition and synthesis is as follows: the operational logic structure of discrete Fourier convolution is divided into two steps. First, channel-independent frequency components are extracted from the time-domain signal using depthwise convolution (equivalent to the multiplication of basis functions in discrete Fourier convolution). Then, these components are linearly weighted and fused using pointwise convolution (equivalent to spectrum synthesis). Finally, the process of extracting the spatiotemporal matrix from the first convolution block of the trained perception model is essentially simulating the process of discrete Fourier convolution and spectrum synthesis.

[0062] See Figure 4 The more structured process of training the perception model involves two parallel processing paths for the spatiotemporal matrix. One path processes the distributed fiber optic sensing signal (spatiotemporal matrix) to be processed through a physical reference branch (Discrete Fourier Transform) to obtain a frequency domain baseline feature representation (physical frequency domain features). The other path involves the first convolutional block of the perception model extracting features from the spatiotemporal matrix to obtain a time-domain learned feature representation (intermediate features), which is then processed by subsequent feature mapping layers to obtain the signal processing output (prediction result). Frequency domain baseline feature representation (physical frequency domain features) and time-domain learned feature representation (intermediate features) are constrained for frequency domain feature consistency. Furthermore, the prediction result and the true label are compared to calculate task-related losses. The resulting loss values ​​are jointly optimized to obtain a total loss value. The parameters of the perception model are updated based on the total loss value until convergence is achieved, resulting in the trained perception model.

[0063] like Figure 5 As shown, this embodiment of the invention provides a specific implementation method for step 202, which specifically includes the following steps: Step 301: Perform preliminary feature extraction on the spatiotemporal matrix through the first convolutional block to obtain intermediate features with frequency domain information.

[0064] The first convolutional block consists of a single convolutional layer, which is a depthwise separable convolution. The first convolutional block is responsible for receiving the spatiotemporal matrix and, guided by the cross-domain distillation mechanism, learning a physical filter similar to the discrete Fourier transform, so that the initially extracted intermediate features have frequency domain related information.

[0065] Step 302: Deep feature extraction is performed on the intermediate features through the second convolutional block to obtain a feature map. The feature map is flattened and input into the recognizer to obtain the prediction result.

[0066] The second convolutional block is responsible for deep feature extraction from intermediate features. Each convolutional layer is followed by a max-pooling layer for spatial downsampling, and an average pooling layer is configured at the end to preserve global context information, thereby enhancing the translation invariance and robustness of the features. The feature map processed by the convolutional block is flattened and input into a recognizer composed of fully connected layers. The recognizer is responsible for mapping the extracted high-dimensional features to specific vibration event category labels.

[0067] The highly lightweight sensing model provided by this invention can be rapidly implemented on FPGA through various hardware description schemes. Its extremely low resource consumption enables it to achieve extremely high throughput under various existing implementation schemes. The solution provided by this invention achieves high computational efficiency through a hardware-friendly lightweight model and ensures strong environmental generalization ability by utilizing physical prior distillation. Compared with the prior art, this solution solves the technical problem that the existing technology is difficult to process in real time for the massive sensing data generated by long-distance fiber optic monitoring in large-scale monitoring applications.

[0068] The sensing model provided by this invention was loaded onto two typical FPGA platforms, and the processing latency, throughput, and resource utilization were evaluated. This demonstrated the enormous potential of combining the lightweight algorithm and hardware acceleration proposed in this solution. On high-end chips, the system exhibited excellent throughput performance, making it suitable for large-scale monitoring over long distances. On resource-constrained industrial chips, it still maintained high throughput and the ability to monitor fiber optic cables over hundreds of kilometers in real time. Ultimately, a low-cost, high-precision, and scalable real-time distributed fiber optic sensing edge computing system was realized.

[0069] Two typical FPGA platforms include a high-performance verification platform (Xilinx ZCU15EG) and a low-end edge computing platform (Xilinx XC7A35T). In the high-performance verification platform, the processing latency of the perception model is as follows: at a clock cycle of 2.24 ns, it takes 8469 clock cycles to complete a single inference, with a total inference latency as low as 0.019 ms. The throughput of the perception model is as follows: it supports processing 13494 samples within a sampling period (e.g., within 256 ms), which enables the system to monitor a 168.7 km long optical fiber in real time with a spatial spacing of 1.25 m. The resource utilization of the perception model is as follows: it consumes 8690 LUTs and 9020 FFs, while the consumption of DSP, Block Random Access Memory (BRAM), and URAM (Ultra Random Access Memory) is zero. Among them, LUTs account for 2.55% and FFs account for 1.32%.

[0070] In low-end edge computing platforms, the processing latency of the perception model is as follows: at a clock cycle of 4.9 ns, a single inference operation requires 8469 clock cycles, with a total inference latency as low as 0.031 ms. The throughput of the perception model supports processing 8246 samples within a sampling period (e.g., within 256 ms), enabling the system to monitor a 103.1 km fiber optic cable in real time with a spatial spacing of 1.25 m. The resource utilization of the perception model is as follows: it consumes 8579 LUTs and 10,068 FFs, with zero consumption of DSP, BRAM, and URAM. Specifically, LUTs account for 20.6% and FFs account for 48.4%.

[0071] To better illustrate the technical details of the present invention, a specific embodiment is provided below, such as... Figure 6 As shown, the present invention can be specifically implemented as an embedded distributed optical fiber vibration detection signal processing system based on a phase-sensitive optical time-domain reflectometer. The system includes a light source emission module, an arbitrary waveform generator, a semiconductor optical amplifier, an erbium-doped fiber amplifier, an optical circulator, a sensing fiber, a photodetector, an analog-to-digital converter, and an embedded processing platform.

[0072] Taking a narrow-linewidth laser as an example, the system employs the following method: an arbitrary waveform generator drives an acousto-optic modulator to convert continuous light into probe pulses with a specific duty cycle, transmitted at a frequency of 1 kHz. An erbium-doped fiber amplifier pre-amplifies the modulated pulses to compensate for linear loss during long-distance fiber transmission. An optical circulator, as the core routing element, guides the amplified pulses from port 1 to port 2 into a 30 km long single-mode sensing fiber. Simultaneously, the circulator guides the backscattered light carrying vibration information returning from the fiber from port 2 to port 3, achieving physical isolation between the probe and receiver light. A photodetector captures the weak optical signal and converts it into an analog electrical signal. Data is acquired via a 250 MHz high-speed analog-to-digital converter, ultimately achieving a spatial spacing of 1.5 meters.

[0073] Example 2: Based on the distributed optical fiber sensor signal real-time processing method provided in the foregoing embodiments, the present invention also provides an apparatus for implementing the above method for distributed optical fiber sensor signal real-time processing, such as... Figure 7 The diagram shown is a schematic representation of the device architecture according to an embodiment of the present invention. The device for real-time processing of distributed fiber optic sensing signals in this embodiment includes one or more processors 21 and a memory 22. Figure 7 Take a processor 21 as an example.

[0074] Processor 21 and memory 22 can be connected via a bus or other means. Figure 7 Taking the example of a connection between China and Israel via a bus.

[0075] The memory 22 serves as a non-volatile computer-readable storage medium for real-time processing of distributed fiber optic sensing signals. It can be used to store non-volatile software programs and non-volatile computer-executable programs, such as the real-time processing method for distributed fiber optic sensing signals in the aforementioned embodiments. The processor 21 executes various functional applications and data processing of the real-time processing device for distributed fiber optic sensing signals by running the non-volatile software programs, instructions, and modules stored in the memory 22, thereby realizing the real-time processing method for distributed fiber optic sensing signals in the aforementioned embodiments.

[0076] Memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, memory 22 may include memory remotely located relative to processor 21, which can be connected to processor 21 via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0077] The program instructions / modules are stored in the memory 22. When executed by one or more processors 21, they perform the real-time processing method for distributed optical fiber sensing signals described in the foregoing embodiments.

[0078] This invention also provides a non-volatile computer storage medium storing computer-executable instructions that are executed by one or more processors, for example... Figure 7 One of the processors 21 enables the one or more processors to execute the distributed optical fiber sensing signal real-time processing method in the foregoing embodiments.

[0079] It is worth noting that the information interaction and execution process between the modules and units in the above-mentioned device and system are based on the same concept as the processing method embodiment of the present invention. For details, please refer to the description in the method embodiment of the present invention, and will not be repeated here.

[0080] Those skilled in the art will understand that all or part of the steps in the various methods of the embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0081] 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, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A physical prior distributed fiber optic sensing real-time method, characterized in that, include: The optical signal transmitted in a long-distance optical fiber is captured, the optical signal is converted into a digital time-series signal, and the digital time-series signal is constructed into a spatiotemporal matrix; A perception model is constructed, and the training of the perception model is guided by extracting frequency domain prior knowledge from optical signal processing, so that the perception model has the ability to extract accurate localization frequency domain information. The accelerator runs the trained perception model and processes the spatiotemporal matrix to identify vibration events along the long-distance optical fiber in the online manner.

2. The real-time distributed fiber optic sensing method with physical prior knowledge according to claim 1, characterized in that, The accelerator includes an FPGA, and the method further includes: The multiplication and accumulation operations in the perception model are replaced with logical shift and addition operations, and hardware acceleration is achieved through a pipelined parallel structure on the FPGA.

3. The real-time distributed fiber optic sensing method with physical prior knowledge according to claim 1, characterized in that, The method further includes: The perception model includes a first convolutional block and a second convolutional block connected in series, and a recognizer connected to the output of the second convolutional block. The first convolutional block includes at least one depth-separable convolutional layer, and the second convolutional block includes at least one depth-separable convolutional layer, a max pooling layer, and an average pooling layer.

4. The real-time distributed fiber optic sensing method with physical prior knowledge according to claim 1, characterized in that, The methods for training the perception model include: Construct a training set, which includes a spatiotemporal matrix generated by optical signal processing and its corresponding real vibration events; The spatiotemporal matrix in the training set is input into the perception model to obtain the prediction result output by the recognizer and the intermediate features extracted by the first convolutional block. Perform a discrete Fourier transform on the spatiotemporal matrix in the training set along the time axis to obtain the physical frequency domain features corresponding to the spatiotemporal matrix; Based on the prediction results, actual vibration events, intermediate features, and physical frequency domain features, the total loss value is calculated according to the defined total loss function. Based on the total loss value, the network parameters are updated using the backpropagation algorithm. The network parameters are updated repeatedly based on the total loss value calculated from the spatiotemporal matrix in the training set until the preset convergence condition is met, and the trained perception model is obtained.

5. The real-time distributed fiber optic sensing method with physical prior knowledge according to claim 4, characterized in that, The total loss function includes matching loss and perceptual loss, wherein the matching loss is used to align the kernel weights of the first convolutional block with the structure of the frequency basis function, and forces the first convolutional block to simulate the frequency selection characteristics of Fourier analysis during gradient descent.

6. The real-time distributed fiber optic sensing method with physical prior knowledge according to claim 5, characterized in that, The total loss function is defined as follows: ； in, Represents the total loss value. Represents the matching loss value. Represents the perceived loss value. This represents the domain weight coefficient.

7. The real-time distributed fiber optic sensing method with physical prior knowledge according to claim 5, characterized in that, The formula for matching loss is as follows: ； in, Represents the matching loss value. This represents the intermediate features extracted from the first convolutional block. It represents the physical frequency domain characteristics.

8. The real-time distributed fiber optic sensing method with physical prior knowledge according to claim 3, characterized in that, The step of inputting the spatiotemporal matrix from the training set into the perceptual model to obtain the prediction result output by the recognizer and the intermediate features extracted by the first convolutional block includes: The first convolutional block is used to perform preliminary feature extraction on the spatiotemporal matrix to obtain intermediate features with frequency domain information; The intermediate features are extracted using a second convolutional block to obtain a feature map. The feature map is then flattened and input into the recognizer to obtain a prediction result.

9. A physical prior distributed fiber optic sensing real-time device, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the processor for performing the physical prior distributed fiber optic sensing real-time method as described in any one of claims 1-8.

10. A non-volatile computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, which are executed by one or more processors to perform the physical prior distributed fiber optic sensing real-time method as described in any one of claims 1-8.

Images