Distributed fiber optic sensing strain measurement method without pre-training

Abstract

This invention relates to the field of fiber optic sensing and detection technology, and more particularly to a distributed fiber optic sensing strain measurement method without pre-training. This method collects reference signals and strain test signals, reconstructs a two-dimensional image through FFT mapping and cross-correlation calculation, skipping the construction of an independent training set and directly using the two-dimensional image generated by the current measurement as the training data source. A sliding window mechanism is used for cyclic prediction until the entire fiber is covered, achieving millimeter-level anomaly localization. Simultaneously, an optimal solution algorithm is used to dynamically optimize the number of model units and the dropout ratio. This method eliminates the traditional solution's reliance on simulated training sets, accelerates hyperparameter optimization, and can meet the real-time processing requirements of OFDR signals.

Description

Technical Field

[0001] The present invention relates to the field of optical fiber sensing and detection technology, and in particular to a distributed optical fiber sensing strain measurement method without pre-training. Background Art

[0002] Distributed fiber-optic sensing technology utilizes the scattering effects of light waves in optical fibers (such as Rayleigh and Brillouin scattering) to enable long-distance, continuous monitoring of physical quantities such as temperature, strain, and vibration. Optical frequency domain reflectometry (OFDR), a representative example of this technology, boasts advantages such as millimeter-level spatial resolution, high sensitivity, and resistance to electromagnetic interference. It is widely used in bridge health monitoring, oil and gas pipeline safety, and aerospace structural inspection.

[0003] Patent CN117629093A discloses a method for improving the spatial resolution of an OFDR shape sensing system based on CNN, which has the following problems:

[0004] 1. The training set is highly dependent, requiring the generation of a large amount of noise-free simulated data in advance to train the model. This requires additional equipment and time costs in actual operation, and cannot respond in real time.

[0005] 2. Lack of local anomaly detection capability: CNN outputs an overall denoising result and cannot locate noise interference at a local location on the fiber, resulting in distortion in strain measurements at high resolution.

[0006] 3. Poor generalization. When the fiber length or sampling rate changes, the training set needs to be regenerated and the network structure adjusted, resulting in low adaptability.

[0007] Patent CN118611052A uses a CNN+GCN+LSTM composite structure to predict distributed photovoltaic power and optimizes hyperparameters using the Whale Algorithm (WOA). However, it has the following problems:

[0008] 1. It is highly dependent on historical data. To build a model, a large amount of historical data on climate factors and photovoltaic power consumption is required as input to predict the true value.

[0009] 2. The computational complexity is too high, the number of composite model parameters is large, and the training and inference delays are significant. It is mainly used for power prediction rather than designed for denoising.

[0010] 3. Optimization efficiency is low. The global search convergence speed of the whale algorithm is slow, and hyperparameter tuning takes a long time.

[0011] To this end, this application designs a distributed fiber optic sensing strain measurement method without pre-training to solve the above problems. Summary of the Invention

[0012] In order to make up for the deficiencies in the prior art, the present invention provides a distributed optical fiber sensing inductive strain measurement method without pre-training.

[0013] A distributed optical fiber sensing strain measurement method without pre-training, characterized by comprising the following steps:

[0014] S1, collects reference signal and test signal;

[0015] S2, the signal is mapped to the range domain by FFT, and the cross-correlation result is calculated according to the window division;

[0016] S3, reconstructing the cross-correlation result into a two-dimensional image, and denoising the two-dimensional image;

[0017] In S4, the denoising process skips the large amount of data training set in the early stage and directly uses the two-dimensional image of S3 as the training data source. The optimal solution algorithm provides hyperparameters for the time series analysis model. The time series analysis model performs cyclic prediction on the local information along the optical fiber to obtain information of the entire sensing fiber.

[0018] S5, hyperparameter optimization is performed through the optimal solution algorithm, and high-resolution strain measurement results are finally output.

[0019] Furthermore, in order to better implement the present invention, the specific steps of the cycle prediction in S4 are:

[0020] S41, selecting the first n points of local information along the optical fiber as an analysis window;

[0021] S42, analyzing whether each point in the window is affected by noise;

[0022] S43, predicting the n+1th point information based on the analysis result;

[0023] S44, the analysis window slides back one point;

[0024] S45, repeat the cycle S42-S44 until the entire optical fiber is covered.

[0025] Furthermore, in order to better implement the present invention, the optimal solution algorithm in S5 is specifically as follows:

[0026] S51, continuing to provide several sets of different model unit numbers and Dropout ratio parameters for the timing analysis model through the optimal solution algorithm;

[0027] S52, the time series analysis model generates several sets of processing results through these several sets of parameters, and uses the mean square error as the evaluation criterion to select the result with the smallest mean square error, which is the best parameter and the best processing result;

[0028] S53, determining whether the maximum number of iterations has been reached, and outputting the result if it has been reached;

[0029] S54, if the maximum number of iterations has not been reached, the optimal solution algorithm will use the current optimal parameters as a basis to generate new parameter combinations in its adjacent area to further approach the global optimum, and then evaluate these new parameters again to continue optimization;

[0030] S55, the entire process is iterated continuously until the maximum iteration condition is met, and the optimal parameters and their corresponding processing results are finally output as the optimal configuration and output of the model in the task.

[0031] The beneficial effects of the present invention are:

[0032] 1. The present invention avoids the traditional reliance on data training sets by directly using a single measured two-dimensional image as training data, thereby improving efficiency and reducing costs.

[0033] 2. The cyclic prediction matrix of the present invention can accurately locate the optical fiber micro-loss point, and the effectiveness of predicting the next sampling point can avoid strain misjudgment caused by multiple peaks / false peaks.

[0034] 3. The present invention adopts the optimal solution algorithm to optimize the number of model units and the dropout ratio, which can significantly improve the robustness of the model; by organizing the training data through a sliding window, it can adapt to different fiber lengths and sampling rates without redesigning the network structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of the present invention;

[0036] Figure 2 This is a flow chart of the traditional method of processing data using neural networks;

[0037] Figure 3 This is the result diagram of the sensing fiber spectrum shift obtained by the traditional processing method;

[0038] Figure 4 This is a diagram of the sensing fiber spectrum shift results obtained by the processing method of the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the accompanying drawings to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in a variety of different configurations.

[0040] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but is merely intended to represent selected embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.

[0041] Figure 1-Figure 4 This is a specific embodiment of the present invention, which is a distributed optical fiber sensing inductive strain measurement method without pre-training.

[0042] like Figure 1 As shown, the traditional neural network processing method requires a large amount of data for iterative training of the network. In this processing method, a large amount of cross-correlation result information is used to construct a data training set, and then the test results are used to compare with the training set to further optimize the test results. This is also a major feature of traditional neural networks, but the disadvantage is that it consumes GPU resources and is complex to adjust parameters. This embodiment only relies on real measurement data to complete training, and does not require a large amount of data for training. It greatly reduces the complexity of data preparation, shortens the system development cycle, and improves the feasibility and promotion of actual engineering applications. By organizing training samples in a sliding window manner, it can flexibly adapt to different sampling rates, fiber lengths and data scale changes without the need to redesign or train the network structure, and has extremely strong input flexibility and environmental adaptability.

[0043] OFDR signals are high-frequency, continuous, and fine-grained data representing physical quantity variations. The complex models of traditional methods cannot meet the real-time processing requirements of OFDR signals. This embodiment utilizes a time series analysis model to directly model the raw data, addressing the high-frequency, continuous, fine-grained variations and long-term dependencies of OFDR signals. This model fully preserves the signal evolution trends and dynamic characteristics, accurately capturing the global evolution process, and effectively addresses the weaknesses of traditional methods in local feature extraction and time series modeling. Combined with an optimal solution algorithm, it can automatically search for and dynamically adjust the optimal parameter combination within the space of key hyperparameters, such as the number of network units and the dropout rate, avoiding the uncertainty and performance fluctuations caused by traditional manual hyperparameter setting. Furthermore, compared to the high computational complexity, slow convergence, and long training cycles of traditional models, this embodiment ensures rapid convergence and improved robustness in diverse environments. This embodiment utilizes a lightweight model design, resulting in a simple and compact time series analysis model structure, fast training and inference speeds, and minimal memory resource usage. It can be deployed on standard CPUs, small GPUs, and even edge computing devices, significantly lowering the system deployment threshold and improving the convenience and universality of practical applications.

[0044] like Figure 2 As shown, the specific method of this embodiment is as follows:

[0045] S1, collects reference signal and test signal;

[0046] S2, the signal is mapped to the range domain by FFT, and the cross-correlation result is calculated according to the window division;

[0047] S3, reconstructing the cross-correlation result into a two-dimensional image, and denoising the two-dimensional image;

[0048] In S4, the denoising process skips the large amount of data training set in the early stage and directly uses the two-dimensional image of S3 as the training data source. The optimal solution algorithm provides hyperparameters for the time series analysis model. The time series analysis model performs cyclic prediction on the local information along the optical fiber to obtain information of the entire sensing fiber. The specific cyclic prediction method is as follows:

[0049] S41, selecting the first n points of local information along the optical fiber as an analysis window;

[0050] S42, analyzing whether each point in the window is affected by noise;

[0051] S43, predicting the n+1th point information based on the analysis result;

[0052] S44, the analysis window slides back one point;

[0053] S45, repeat the cycle S42-S44 until the entire optical fiber is covered.

[0054] S5, performs hyperparameter optimization through the optimal solution algorithm and finally outputs high-resolution strain measurement results;

[0055] The optimal solution algorithm is specifically as follows:

[0056] S51, continuing to provide several sets of different model unit numbers and Dropout ratio parameters for the timing analysis model through the optimal solution algorithm;

[0057] S52, the time series analysis model generates several sets of processing results through these several sets of parameters, and uses the mean square error as the evaluation criterion to select the result with the smallest mean square error, which is the best parameter and the best processing result;

[0058] S53, determining whether the maximum number of iterations has been reached, and outputting the result if it has been reached;

[0059] S54, if the maximum number of iterations has not been reached, the optimal solution algorithm will use the current optimal parameters as a basis to generate new parameter combinations in its adjacent area to further approach the global optimum, and then evaluate these new parameters again to continue optimization;

[0060] S55, the entire process is iterated continuously until the maximum iteration condition is met, and the optimal parameters and their corresponding processing results are finally output as the optimal configuration and output of the model in the task.

[0061] from Figure 2 As can be seen, compared to traditional data processing methods, this embodiment avoids the drawbacks associated with constructing a training set using large amounts of training data. Instead, it employs a combined global and local search strategy within the resulting cross-correlation information. For example, it selects the first 20 points along the fiber to analyze whether points 0-20 are affected by noise, and then predicts whether the information at point 21 is correct. This process is repeated repeatedly to obtain information for the entire sensing fiber. This processing approach significantly improves processing efficiency and allows strain measurements with higher spatial resolution to be achieved using this embodiment.

[0062] In addition, from Figure 1 Figure 2 From the flowchart, the traditional neural network processing method focuses more on the overall network of the structure, relying only on the neural network to process and denoise the signal, without making reasonable use of the data evaluation mechanism to reduce the difficulty of building the model, which leads to the complexity of model construction. This embodiment can effectively reduce the complexity of the model. In addition, this embodiment can adaptively adjust the parameters of the time series analysis model through the optimal solution algorithm, avoiding manual parameter adjustment, and has better prediction accuracy. It can be applied to different noise environments, while the traditional neural network processing method uses fixed parameters, has no adaptability, and requires manual parameter adjustment.

[0063] like Figure 3 and Figure 4 It can be seen from the comparison that, after experiments, at a spatial resolution of 1.5 mm, the traditional processing method has difficulty in distinguishing different strains and is very likely to cause erroneous analysis, while the method of this embodiment can well analyze the strain results.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not limiting. Other modifications or equivalent substitutions made to the technical solution of the present invention by ordinary technicians in this field should be included in the scope of the claims of the present invention as long as they do not depart from the spirit and scope of the technical solution of the present invention.

Claims

1. A distributed optical fiber sensing strain measurement method without pre-training, characterized in that: The following steps are involved: S1, collects reference signal and test signal; S2, the signal is mapped to the range domain by FFT, and the cross-correlation result is calculated according to the window division; S3, reconstructing the cross-correlation result into a two-dimensional image, and denoising the two-dimensional image; In S4, the denoising process relies solely on real measurement data, skipping the large amount of data training sets in the early stages. It directly uses the two-dimensional image from S3 as the training data source and uses the optimal solution algorithm to provide hyperparameters for the time series analysis model. The time series analysis model performs cyclic prediction on the local information along the optical fiber to obtain information about the entire sensing fiber. The specific steps of cyclic prediction are as follows: S41, selecting the first n points of local information along the optical fiber as an analysis window; S42 analyzes whether each point in the window is affected by noise. The model automatically models the inherent pattern by learning the changing trends of n-point information windows. Even if some points are jumping noise, the overall trend can be captured, making the generated n+1th point smoother. S43, predicting the n+1th point information based on the analysis result; S44, the analysis window slides back one point; S45, repeating the cycle S42-S44 until the entire optical fiber is covered; S5, hyperparameter optimization is performed through the optimal solution algorithm, and high-resolution strain measurement results are finally output.

2. The distributed optical fiber sensing inductive strain measurement method without pre-training according to claim 1, characterized in that: The optimal solution algorithm in S5 is specifically: S51, continuing to provide several sets of different model unit numbers and Dropout ratio parameters for the timing analysis model through the optimal solution algorithm; S52, the time series analysis model generates several sets of processing results through these several sets of parameters, and uses the mean square error as the evaluation criterion to select the result with the smallest mean square error, which is the best parameter and the best processing result; S53, determining whether the maximum number of iterations has been reached, and outputting the result if it has been reached; S54, if the maximum number of iterations has not been reached, the optimal solution algorithm will use the current optimal parameters as a basis to generate new parameter combinations in its adjacent area to further approach the global optimum, and then evaluate these new parameters again to continue optimization; S55, the entire process is iterated continuously until the maximum iteration condition is met, and the optimal parameters and their corresponding processing results are finally output as the optimal configuration and output of the model.

Images