FPI sensing method and system based on chirped pulse and graph convolutional neural network

Abstract

The invention discloses an FPI sensing system based on chirp pulse detection and a graph convolutional neural network, and the system comprises a chirp generation module which is used for generating chirp pulse incident light and splitting the chirp pulse incident light into a first chirp pulse and a second chirp pulse; the physical quantity detection module is used for respectively inputting the first chirped pulse and the second chirped pulse into the reference FPI and the sensing FPI for physical quantity detection to obtain a first interference signal and a second interference signal and coupling the first interference signal and the second interference signal into a third interference signal; the signal acquisition and processing module is used for acquiring a third interference signal, preprocessing the third interference signal and then constructing an original time-frequency signal diagram; and the signal demodulation module is used for predicting the preprocessed original time-frequency signal graph based on the time-frequency fusion graph convolutional network to obtain a frequency difference after sensing FPI physical quantity detection, demodulation is performed based on a mapping relation between the cavity length change and the frequency difference to obtain sensing FPI cavity length change, and a physical quantity detection result is determined based on the sensing FPI cavity length change.

Description

Technical Field

[0001] The present invention relates to the fields of optical computing and optical fiber sensing, and in particular to an FPI sensing method and system based on chirped pulses and graph convolutional neural networks. Background Art

[0002] Sensors are one of the three pillar industries in today's society, and almost all industries cannot do without their applications. As the use environment becomes more and more complex, more and more stringent requirements are placed on the performance of sensors, among which temperature and pressure are important monitoring and measurement items. Optical fiber sensors play an increasingly important role in the sensor industry due to their resistance to high temperature and high pressure, corrosion resistance, long-distance transmission, anti-electromagnetic interference, and high stability in harsh environments.

[0003] As an interferometric fiber optic sensor, the temperature and pressure sensing technology of Fabry-Perot Interferometer (FPI) is a high-precision sensing and measurement technology. The reflected light from the two reflecting surfaces in the FPI sensor cavity interferes due to the optical path difference. When external parameters such as temperature and pressure act on the FPI cavity, its cavity length will change accordingly, causing the reflection spectrum of the FPI to change. The purpose of sensing is achieved through light intensity demodulation, phase demodulation, wavelength demodulation and other technologies.

[0004] However, as a sensor that uses the propagation characteristics of light as a sensing signal, the sensing ability of the FPI sensor is affected by many factors, such as the uniformity of components, the stability of the detection light, and external interference, etc. These factors restrict the further improvement of its sensitivity and the cost control of the system. Among the current FPI cavity sensor demodulation algorithms, although the intensity demodulation is simple in principle and easy to implement, the demodulation accuracy is low; although the phase demodulation has the characteristics of high demodulation accuracy and fast demodulation speed, there are problems such as unstable output phase signal and insufficient system anti-interference ability; the wavelength demodulation algorithm is complex and cannot meet the purpose of fast demodulation. In addition, each method requires calibration of different sensors to achieve quantitative measurement of the measurement to be measured, which also increases the cost of the sensor.

[0005] The specification with publication number CN 115638894 A discloses a sensitivity-enhanced composite multi-cavity FPI temperature sensor, preparation method and application. In this sensor structure, a composite multi-cavity FPI is formed. When the temperature rises, different interference cavities produce opposite cavity length changes, that is, different interference cavities have anti-phase characteristics in response to temperature. The demodulation sensitivity is improved by calculating the phase difference between the air cavity FPI and the air PDMS cavity FPI. The prior art adopts a differential phase demodulation method, which is highly complex. At the same time, the sensing resolution is limited by the signal-to-noise ratio of the optical fiber to be measured, and there is a problem of insufficient system anti-interference ability. Summary of the invention

[0006] The purpose of the present invention is to provide a FPI sensing method and system based on chirped pulses and graph convolutional neural networks, reduce the complexity of the demodulation method, avoid the problem that the traditional demodulation method is limited by the swept light source or spectrometer, and improve the demodulation capability and detection accuracy of the system.

[0007] The present invention provides a FPI sensing system based on chirped pulse detection and graph convolutional neural network, which is characterized by comprising:

[0008] A chirp generation module, which is used to generate a chirped pulse incident light and split it into a first chirped pulse and a second chirped pulse;

[0009] A physical quantity detection module, which is used to input the first chirped pulse and the second chirped pulse into the reference FPI and the sensing FPI respectively for physical quantity detection, to obtain the first interference signal and the second interference signal and couple them into a third interference signal;

[0010] A signal acquisition and processing module, which is used to acquire the third interference signal and construct an original time-frequency signal diagram after preprocessing;

[0011] The signal demodulation module is used to predict the preprocessed original time-frequency signal graph based on the time-frequency fusion graph convolutional network, obtain the frequency difference after the sensing FPI physical quantity detection, and demodulate based on the mapping relationship between the cavity length change and the frequency difference to obtain the sensing FPI cavity length change, and then determine the physical quantity detection result based on the sensing FPI cavity length change.

[0012] Further, the physical quantity detection module includes a first circulator, a second circulator, a reference FPI, a sensing FPI and an optical coupling unit;

[0013] The first chirped pulse is input into the reference FPI after passing through the first circulator, and the first interference signal formed after being reflected by the reference FPI is input into the optical coupling unit through the first circulator. The second chirped pulse is input into the sensing FPI after passing through the second circulator, and the second interference signal formed after being reflected by the sensing FPI is input into the optical coupling unit through the second circulator. The optical coupling unit couples the two input interference signals and outputs the third interference signal to the signal acquisition and processing module.

[0014] The reference FPI adopts a solid cavity structure with calibrated parameters; the sensing FPI includes at least one cavity with variable length, and the length of the cavity changes with the change of the physical quantity to be measured.

[0015] Furthermore, in the signal acquisition and processing module, the third interference signal is acquired and preprocessed to construct an original time-frequency signal graph, including:

[0016] After the collected third interference signal is photoelectrically converted, it is filtered and denoised, and then the interference signal after denoising is sliced, and the obtained slice vector is used as the node, the slice mode is used as the edge, and the strength of the correlation between the nodes is used as the edge weight to construct the original time-frequency signal graph of the interference signal, where the slicing mode includes pixel overlap and time window.

[0017] Furthermore, the time-frequency fusion graph convolutional network includes a graph autoencoder and a time-frequency feature extraction module, as well as a frequency difference prediction module; the graph autoencoder module is used to reconstruct the graph based on the input original time-frequency signal graph to obtain a reconstructed time-frequency signal graph; the time-frequency feature extraction module is used to extract time-frequency features from the reconstructed time-frequency signal graph; the frequency domain difference prediction module is used to predict the frequency difference after each cavity in the sensor FPI changes based on the time-frequency features.

[0018] Furthermore, the graph autoencoder includes a graph encoder, a graph decoder, and a multi-head attention module. The slice vector sequence is used as the input of the multi-head attention module to calculate the attention weight, and the attention weight of each slice vector is introduced into the graph encoder and the graph decoder. The input original time-frequency signal graph encodes each node in the graph encoder based on the introduced attention weight to obtain the encoded feature, and the encoded feature of each node is decoded in the graph decoder based on the introduced attention weight to reconstruct the graph to obtain the reconstructed time-frequency signal graph.

[0019] The time-frequency feature extraction module includes at least one convolutional neural network and at least two graph convolutional networks connected in sequence, wherein the convolutional neural network is used to extract the timing features in the reconstructed time-frequency signal graph, and the graph convolutional network is used to extract the time-frequency features of each node in the reconstructed time-frequency signal graph based on the extracted timing features.

[0020] The frequency difference prediction module uses a fully connected layer to predict the frequency difference at different times based on the time-frequency characteristics of each node.

[0021] The time-frequency fusion graph convolutional network is trained to optimize parameters before being applied. The loss function used in the training process includes the reconstruction loss between the original time-frequency signal graph and the reconstructed time-frequency signal graph, and also includes a semi-supervised classification loss constructed based on the predicted frequency difference and the actual frequency difference.

[0022] The training process of the time-frequency fusion graph convolutional network includes:

[0023] Prepare a dataset of original time-frequency signal graphs, and label a small part of the original time-frequency signal graphs with the real frequency difference to form a labeled dataset, and the remaining large part of the original time-frequency signal graphs constitute an unlabeled dataset;

[0024] The time-frequency fusion graph convolutional network is trained for the first time using a labeled data set. The loss function used in the first training includes a reconstruction loss and a supervised classification loss based on the real frequency difference and the predicted frequency difference, thereby obtaining a first training model.

[0025] The first training model is used to predict the unlabeled data set, and the prediction results are used as pseudo labels to obtain a temporary unlabeled data set with pseudo labels. A part of pseudo labels are selected from the temporary unlabeled data set according to a set threshold as credible pseudo labels to form a pseudo label data set; the first training model is trained for a second time using the pseudo label data set, and the loss function used in the second training includes reconstruction loss and semi-supervised classification loss based on the frequency difference between the pseudo label and the prediction, to obtain a second training model;

[0026] Fine-tuning the second training model using the labeled data set, wherein the loss function used in the fine-tuning includes a reconstruction loss and a supervised classification loss based on the true frequency difference and the predicted frequency difference, to obtain a third training model;

[0027] The second and third training processes are repeated until the prediction results of the model reach the set prediction accuracy, completing the parameter optimization of the time-frequency fusion graph convolutional network.

[0028] Furthermore, the mapping relationship between the cavity length change and the frequency difference is expressed as:

[0029]

[0030] Where Δf is the fixed frequency difference of the interference signal of the reference FPI, L is the fixed cavity length of the reference FPI; i=1,2,... represents the i-th cavity in the sensing FPI, Δf i is the first variable frequency difference of the i-th cavity in the sensing FPI before the measured physical quantity changes, Δf i ' is the second variable frequency difference of the i-th cavity in the sensing FPI after the measured physical quantity changes, ΔL i It represents the change in the length of the i-th cavity of the sensing FPI before and after the physical quantity to be measured changes.

[0031] The present invention also provides a FPI sensing method based on chirped pulse detection and graph convolutional neural network. The method is implemented by the FPI sensing system based on chirped pulse detection and graph convolutional neural network, and comprises the following steps:

[0032] Step 1: Keep the physical quantity to be measured of the reference FPI unchanged, change the physical quantity to be measured of the sensing FPI multiple times, and obtain multiple interference signals containing cavity length information and frequency information through the chirp generation module and the physical quantity detection module;

[0033] Step 2: Preprocess all interference signals through the signal acquisition and processing module and construct the original time-frequency signal graph;

[0034] Step 3: The signal demodulation module predicts the original time-frequency signal graph based on the time-frequency fusion graph convolutional network to obtain the frequency difference after the sensor FPI physical quantity detection, and demodulates the sensor FPI cavity length change based on the mapping relationship between the sensor FPI cavity length change and the frequency difference to obtain the sensor FPI cavity length change, and then determines the physical quantity detection result based on the sensor FPI cavity length change.

[0035] Compared with the prior art, the FPI sensing system and method based on chirped pulse detection and graph convolutional neural network proposed in the present invention has the following beneficial effects:

[0036] (1) The use of chirped pulses as detection light avoids the problem of using a swept-frequency light source or spectrometer in traditional demodulation methods, which can significantly reduce the cost of the sensing system.

[0037] (2) By using an FPI with known parameters as a reference and analyzing the mapping relationship between the optical path difference and the frequency difference between the detection light and the reflected light by referring to the cavity length of the FPI, the demodulation of the sensing FPI cavity length can be completed. The demodulation is simple, with few input parameters and low complexity.

[0038] (3) The time-frequency graph is reconstructed using a graph autoencoder based on a multi-head attention mechanism, which improves the model's ability to extract features from graph data and enhances its ability to process time-frequency signals.

[0039] (4) Semi-supervised training and parameter optimization are performed on the time-frequency fusion graph convolutional network model, which not only reduces the dependence on labeled data and improves sensing efficiency, but also enhances the model's generalization ability, making it suitable for a variety of application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Schematic diagram of an FPI sensing system based on chirped pulse detection and graph convolutional neural network in an embodiment of the present invention;

[0041] Figure 2 Schematic diagram of the internal structure of the sensing FPI and the reference FPI in an embodiment of the present invention, wherein: Figure 2 (a) shows the longitudinal section of the sensing FPI, Figure 2 (b) represents the longitudinal section of the reference FPI;

[0042] Figure 3 Schematic diagram of time-frequency signals of interference signals in an embodiment of the present invention;

[0043] Figure 4 Flow chart of interference signal preprocessing and slicing in an embodiment of the present invention;

[0044] Figure 5 It is a structural diagram of a time-frequency fusion graph convolutional network in an embodiment of the present invention;

[0045] Figure 6 It is a structural diagram of the autoencoder in an embodiment of the present invention;

[0046] Figure 7 It is a training flow chart of the time-frequency fusion graph convolutional network model in an embodiment of the present invention;

[0047] Figure 8 Flow chart of an FPI sensing method based on chirped pulse detection and graph convolutional neural network in an embodiment of the present invention;

[0048] The markings in the attached drawings are:

[0049] 10. Chirp generation module; 11. Chirp pulse emission unit; 12. Spectral splitting unit;

[0050] 20. Physical quantity detection module; 21. First circulator; 22. Second circulator; 23. Reference FPI; 24. Sensing FPI; 25. Optical coupling unit;

[0051] 30. Signal acquisition and processing module; 31. Photoelectric detection unit; 32. Signal processing unit;

[0052] 40. signal demodulation module; 41. user terminal unit; 42. demodulation algorithm unit;

[0053] 231, reference FPI cavity; 232, single mode optical fiber;

[0054] 241, sensing FPI primary cavity; 242, sensing FPI secondary cavity;

[0055] R 11 , reference to the first reflection surface of FPI; R 12 , the second reflection surface of the reference FPI; L, the fixed cavity length of the reference FPI;

[0056] R 21 , the first reflective surface of the sensing FPI; R 22 , the second reflective surface of the sensing FPI; R 23 , the third reflective surface of the sensing FPI; L 1 , sensing the variable cavity length of the FPI primary cavity; L 2 , sensing the variable cavity length of the FPI secondary cavity. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0058] The present invention adopts an ultrashort chirped light pulse as a detection signal, introduces a reference FPI with known and calibrated parameters, and realizes demodulation of the unknown cavity length of the sensor FPI to be measured by comparing the return interference signal of the sensor FPI structure to be measured excited by the chirped pulse and the FPI with known parameters. The change of the cavity length can directly reflect the change of the physical quantity, so the measurement of the physical quantity can be further realized.

[0059] At the same time, in order to improve the demodulation accuracy and automatic demodulation capability of the sensing system, the present invention also designs a time-frequency fusion graph convolutional neural network (TF-GCN) based on graph autoencoder and multi-head attention mechanism for multi-FP cavity signal demodulation.

[0060] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0061] A FPI sensing system based on chirped pulse detection and graph convolutional neural network, such as Figure 1 As shown, including:

[0062] The chirp generation module 10 comprises a chirped pulse emitting unit 11 and a light splitting unit 12, which are used to generate chirped pulse incident light and split it into a first chirped pulse and a second chirped pulse.

[0063] The physical quantity detection module 20 includes a first circulator 21, a second circulator 22, a reference FPI 23, a sensing FPI 24 and an optical coupling unit 25, which is used to input the first chirped pulse and the second chirped pulse into the reference FPI 23 and the sensing FPI 24 respectively for physical quantity detection, and obtain the first interference signal and the second interference signal and couple them into a third interference signal through the optical coupling unit 25.

[0064] The signal acquisition and processing module 30 includes a photoelectric detection unit 31 and a signal processing unit 32, which are used to acquire the third interference signal and construct an original time-frequency signal diagram after preprocessing.

[0065] The signal demodulation module 40 includes a user terminal unit 41 and a demodulation algorithm unit 42, wherein the user terminal unit 41 is used for user-oriented interactive operations, and the demodulation algorithm unit 42 is used to predict the preprocessed original time-frequency signal graph based on the time-frequency fusion graph convolution network, obtain the frequency difference after the sensing FPI physical quantity detection, and demodulate based on the mapping relationship between the cavity length change and the frequency difference to obtain the sensing FPI cavity length change, and then determine the physical quantity detection result based on the sensing FPI cavity length change.

[0066] In this embodiment, a beam of ultra-short linear chirped pulses is generated by a chirped pulse emitting unit 11, which is connected to a splitting unit 12 using a 50:50 splitting unit through a connecting optical fiber, and is split into a first chirped pulse and a second chirped pulse; two connecting optical fibers are led out from the other end of the 50:50 splitting unit, which are used to be connected to a first circulator 21 and a second circulator 22 of a multi-physical quantity detection module 20.

[0067] The optical coupling unit 25 uses a 1×2 optical coupler. Figure 1 The solid connecting lines in the physical quantity detection module 20 represent connecting optical fibers. The first circulator 21 includes three interfaces, which are respectively connected to the 50:50 splitting unit, the reference FPI 23, and the 1×2 optical coupler through connecting optical fibers; the second circulator 22 includes three interfaces, which are respectively connected to the 50:50 splitting unit, the sensing FPI, and the 1×2 optical coupler through connecting optical fibers; the 1×2 optical coupler leads out the connecting optical fiber to the signal acquisition and processing module.

[0068] The single FP cavity structure of the reference FPI 23 used in this embodiment is as follows: Figure 2 As shown in (a), it includes a cavity 231, two reflective surfaces R 11 and R 12 And the single-mode optical fiber 232 adopts a solid cavity structure with parameters calibrated, and the cavity length L is a known quantity.

[0069] The sensing FPI 24 can be a single-cavity structure or can be composed of multiple FP cavities in cascade, where the length of the i-th FP cavity is defined as L i , i=1,2,3,..., cavity length L i Not determined; the Figure 2 (b) shows a dual-cavity cascade FPI formed by three interfaces as a sensing FPI 24, including a primary cavity 241 and a secondary cavity 242, and three reflective surfaces R 21 , R 22 and R 23 Among them, the first-stage cavity 241 and the second-stage cavity 242 are both cavities with variable cavity lengths, and the first-stage cavity length L 1 and the secondary cavity length L 2It changes with the change of the physical quantity to be measured. Since the refractive index of the materials on both sides of each interface is different, the light is reflected by each interface during the propagation process, and each pair of interfaces can form an FPI.

[0070] Since the incident light is an ultra-short linear chirped pulse, different frequency components therein will reach the reflecting surface inside the FPI at different times and be transmitted again at slightly different time intervals. By using chirped pulses as detection light, a narrow-linewidth detection light source with high spectral purity, long coherence length, and low phase noise, or even a single-wavelength optical signal carrier chirped pulse, can be used to significantly reduce noise interference and improve detection accuracy. This avoids the problem of traditional demodulation methods requiring the use of swept-frequency light sources or spectrometers, and can significantly reduce the cost of the sensing system.

[0071] In this embodiment, the process of obtaining the interference signal by performing physical quantity detection using chirped pulses is specifically as follows:

[0072] Figure 3 is a schematic diagram of the time-frequency signal of the interference signal, and the width of the chirped pulse is T=t 3 -t 1 =t 7 -t 4 , divided into two beams of identical first chirped pulses and second chirped pulses.

[0073] The first chirped pulse is input to the reference FPI 23 after passing through the first circulator 21. 1 Reaching the reflecting surface R 11 Partial reflection occurs at t and part of the transmission enters the cavity 231. 2 Arrives at the reflecting surface R 12 Partial reflection and partial transmission occur again, at t 3 Moment reflection surface R 11 The received pulse ends at t 10 Moment reflection surface R 12 The received pulse ends, and finally the first interference signal carrying the fixed cavity length information L of the cavity 231 is obtained; the time difference of the first chirped pulse between the two reflection surfaces is ΔT=t 2 -t 1 , and since the cavity length L of the reference FPI is a constant value, the frequency difference of the cavity 231 is fixed to Δf, the time domain signal of the first interference signal is a sinusoidal signal with a frequency of Δf, and the first interference signal is input to the optical coupling unit 25 through the first circulator 22.

[0074] The second chirped pulse is input to the sensing FPI 24 after passing through the second circulator 22. 4 Arrives at the reflecting surface R 21 Partial reflection and partial transmission into the primary cavity 241 are generated, and then at t5 Arrives at the reflecting surface R 22 Partial reflection and partial transmission into the secondary cavity 242 are generated, and at t 6 Arrives at the reflecting surface R 23 Partial reflection and partial transmission occur at t 7 Moment reflection surface R 21 The received pulse ends at t 8 Moment reflection surface R 22 The received pulse ends at t 9 Moment reflection surface R 23 The received pulse ends, and finally the variable cavity length information L of the first-stage cavity 241 is obtained. 1 and the variable cavity length information L of the secondary cavity 242 2 The second interference signal of the second chirped pulse is 21 and the reflective surface R 22 The time difference between them is ΔT 1 =t 5 -t 4 , on the reflecting surface R 22 and the reflective surface R 23 The time difference between them is ΔT 2 =t 6 -t 5 , on the reflecting surface R 21 and the reflective surface R 23 The time difference between them is ΔT 3 =t 6 -t 4 The first variable frequency difference of the first cavity 241 is Δf 1 , the first variable frequency difference of the secondary cavity 242 is Δf 2 , reflecting surface R 21 and the reflective surface R 23 The first variable frequency difference between 3 The second interference signal is then input into the optical coupling unit 25 through the second circulator 22.

[0075] The input first interference signal and the second interference signal are coupled through a 1×2 optical coupler, and then the third interference signal is output to the signal acquisition and processing module 30 .

[0076] The photoelectric detection unit 31 in the signal acquisition and processing module 30 receives the third interference signal transmitted from the 1×2 optical coupler through the connecting optical fiber. Figure 4The processing process from the third interference signal to the slice vector is demonstrated. After the third interference signal is converted into an electrical signal, it is transmitted to the signal processing unit 32 through a connecting cable for analog-to-digital conversion, filtering and noise reduction processing, and then the interference signal after noise reduction processing is sliced ​​to generate a time-frequency information matrix of the interference signal, and the time-frequency information matrix is ​​passed to the signal demodulation module 40; wherein, the filtering and noise reduction processing can use wavelet transform or Fourier transform, and in this embodiment, Fourier transform is used for filtering and noise reduction processing.

[0077] The obtained slice vector is used as the node, the slice mode is used as the edge, and the strength of the correlation between the nodes is used as the edge weight to construct the original time-frequency signal graph of the interference signal, which is expressed as:

[0078]

[0079] Among them, x i represents the i-th node, X represents the node matrix composed of N nodes; e ij represents the edge between the i-th node and the j-th node, E represents the edge matrix composed of all edges. In the original graph, the edge is represented as an adjacency matrix. When there is pixel overlap between two nodes, there is an edge; w ij Represents edge e ij The weight of , W represents the weight matrix composed of all weights. In the original graph, the edge weight is calculated by cosine similarity to indicate the strength of the edge; the slicing method includes pixel overlap and time window.

[0080] Due to the different physical structures and material properties of the FP cavities at each stage of the sensor FPI, and the time-frequency signals of the reflected signals obtained after each real-time detection are not constant, the TF-GCN model is designed to extract the mapping relationship between the frequency that changes with time and the cavity length of each stage of the sensor FPI from the time-frequency signal.

[0081] In order to improve the generalization ability of the TF-GCN model based on the sensor system in the present invention for real-time demodulation of various detection scenarios, learn more robust feature representations in time-frequency signals, and improve the model's ability to extract features from graph data, the original time-frequency graph is reconstructed in the graph autoencoder. Through graph reconstruction, not only can the performance of TF-GCN be optimized, the use of computing resources in the signal demodulation module 40 can be improved, and the use of memory can be optimized, but also the efficiency and accuracy of model training and frequency feature extraction can be improved, and the structure and features in the original time-frequency signal graph can be more effectively extracted and utilized.

[0082] like Figure 5As shown, the TF-GCN includes a graph autoencoder, a time-frequency feature extraction module, and a frequency difference prediction module; the graph autoencoder module is used to reconstruct the graph based on the input original time-frequency signal graph to obtain a reconstructed time-frequency signal graph; the time-frequency feature extraction module is used to extract time-frequency features from the reconstructed time-frequency signal graph; the frequency domain difference prediction module is used to predict the frequency difference after each cavity changes in the sensing FPI based on the time-frequency features.

[0083] The original time-frequency signal graph is input into TF-GCN, which can be expressed as TF-GCN (X, E, W). The feature extraction of interference signals through TF-GCN can greatly improve the efficiency of GPU usage while improving parallel computing capabilities. Combined with the weight adjustment of multi-head attention and the reconstruction of the graph by the graph encoder, the temporal dependency and spatial correlation of subsequent modules are enhanced, which enables effective learning and feature extraction, and realizes the demodulation of interference signals.

[0084] Figure 6 The process of constructing the slice vector into the original time-frequency signal graph through the graph autoencoder is demonstrated. The graph autoencoder includes a graph encoder, a graph decoder, and a multi-head attention module, which is used to reconstruct the original time-frequency graph and remove the low-weight edges between nodes to further reduce the noise of the interference signal.

[0085] The multi-head attention module uses the slice vector sequence as the input of the multi-head attention module to calculate the attention weight, and introduces the attention weight of each slice vector into the graph encoder and the graph decoder. The input original time-frequency signal graph is encoded by embedding each node in the low-dimensional vector space based on the introduced attention weight in the graph encoder to obtain the encoded feature. The encoded feature of each node is decoded in the graph decoder based on the introduced attention weight to reconstruct the graph and obtain the reconstructed time-frequency signal graph.

[0086] In the graph autoencoder, the original node matrix X and the original adjacency matrix A of the original time-frequency signal graph are used as input, and the encoding feature Z=f(A,X) is obtained by the graph encoder, and then the reconstructed node matrix X and the reconstructed adjacency matrix A are obtained by the graph decoder;

[0087] The multi-head attention module performs self-attention learning and weight allocation on the nodes divided into p parts. The mth head after division can be expressed as:

[0088] head(m)=Attention(Q m ,K m ,V m ),m=1,2,...,p

[0089] in, is the query matrix of the mth head, which is used to find other elements related to itself from the segment sequence. Q represents the original query matrix. is the key matrix of the mth head, which is used to characterize the feature information of each position in the segment sequence. K represents the original key matrix. is the value matrix of the mth head, used to calculate the output after attention, and V represents the original value matrix; Q m , K m 、V m A learnable weight matrix is ​​used to project the original Q, K, V into the subspace specific to the mth head;

[0090] The self-attention learning process of the mth head can be expressed as:

[0091]

[0092] Among them, L m is the scaling factor related to the mth head, by adjusting L m Make the graph autoencoder converge;

[0093] All heads are merged, expressed as:

[0094] MultiHead(Q,K,V)=Concat{head(1),head(2),...,head(p)}W O

[0095] Among them, W O Is a linear transformation matrix used to map the fused high-dimensional vector back to the original dimension;

[0096] The purpose of reconstructing the original time-frequency graph signal graph by using the graph encoder based on the multi-head attention mechanism is to improve the generalization ability of the TF-GCN model based on the sensor system to perform real-time demodulation in various detection scenarios, and to learn more robust feature representations in time-frequency signals; at the same time, optimize the performance and memory usage of TF-GCN, improve the use of computing resources in the signal demodulation module 40, improve the efficiency and accuracy of model training and frequency feature extraction, and effectively extract and utilize the structure and features of the original time-frequency signal graph.

[0097] The output of the graph autoencoder is used as the input of the time-frequency feature extraction module, such as Figure 5 As shown, in order to capture the frequency correlation and time dependence of the reconstructed time-frequency signal graph, this embodiment designs the time-frequency feature extraction module as two convolutional neural networks and three graph convolutional networks connected in sequence;

[0098] According to the reconstructed node matrix X and the reconstructed adjacency matrix obtained by the graph decoder, the corresponding reconstruction degree matrix D is obtained, and together they form a reconstructed time-frequency signal graph as the input of the time-frequency feature extraction module;

[0099] The time series features of the nodes in the reconstructed time-frequency signal graph are extracted through the convolutional neural network. The node features of the lth layer of the convolutional neural network can be updated by the following formula:

[0100]

[0101] in, represents (the edge weight matrix of the lth layer after update), Represents (output of layer l-1);

[0102] On the basis of extracting the time series features, the time-frequency features of each node in the time-frequency signal graph are extracted and reconstructed through the graph convolution network; among them, the Laplace matrix L is constructed r It is used to capture the structural information of the reconstructed time-frequency signal graph, expressed as:

[0103]

[0104] Where U is L r The eigenvector matrix of L r The eigenvalue diagonal matrix of T is a unit diagonal matrix. Based on the above formula, we further perform graph convolution operations. To this end, we first define a convolution kernel g(θ). The scope of each convolution operation is expected to be in the area near the central node. Then g can be regarded as a function g(L) about the Laplace matrix L. Each convolution is equivalent to transferring the information of adjacent nodes. The convolution operation is performed through Since the graph G of N nodes has a very high time complexity, the Chebyshev polynomial is used to parameterize the convolution kernel g(θ). Therefore, the node features of the sth layer of the graph convolution network can be updated by the following formula:

[0105]

[0106] in:

[0107] represents the output of the sth layer, Z s-1 represents the output of the s-1th layer,

[0108] σ represents the activation function; represents Chebyshev polynomials;

[0109] U represents the eigenvector matrix of the Laplacian matrix of the reconstructed time-frequency graph, U T is transposed;

[0110] k<<number of nodes n, indicating the kth order neighbor;

[0111] θ′ k is the angle representation of the Chebyshev polynomial, where k∈(1,K)

[0112] represents the kth order Chebyshev polynomial, in which x = cos(θ′ k ), where T 0 (x) = 1, T 0 (x) = x, ..., T k (x) = 2xT k-1 (x)-T k-2 (x), x∈[-1, 1];

[0113]

[0114] The output of the time-frequency feature extraction module is output through a frequency difference prediction module, and the frequency difference prediction module uses a fully connected layer to predict the frequency difference at different times based on the time-frequency features of each node.

[0115] The time-frequency fusion graph convolutional network is trained to optimize parameters before being applied. The loss function used in the training process includes the reconstruction loss between the original time-frequency signal graph and the reconstructed time-frequency signal graph. The loss function also includes a semi-supervised classification loss based on the predicted frequency difference and the actual frequency difference. Need to be based on and For the multi-head attention module Make dynamic adjustments.

[0116] The reconstruction loss and semi-supervised classification loss are intended to make the node representation learned by the graph autoencoder reflect the true structure of the original time-frequency signal graph as accurately as possible, improve the robustness and stability of the system, reduce the manual adjustment of hyperparameters during the learning process, and improve the model's ability to capture the interactivity between variables in the reference FPI and sensor FPI datasets.

[0117] The training process of the time-frequency fusion graph convolutional network is as follows Figure 7 As shown:

[0118] Prepare an original time-frequency signal graph dataset, and label a small part of the original time-frequency signal graph with a real frequency difference to form a labeled dataset, and the remaining large part of the original time-frequency signal graph constitutes an unlabeled dataset. In this embodiment, the ratio of the labeled dataset to the unlabeled dataset is 1:9;

[0119] The time-frequency fusion graph convolutional network is trained for the first time using a labeled data set. The loss function used in the first training includes a reconstruction loss and a supervised classification loss based on the real frequency difference and the predicted frequency difference, thereby obtaining a first training model.

[0120] The first training model is used to predict the unlabeled data set, and the prediction results are used as pseudo labels to obtain a temporary unlabeled data set with pseudo labels. A part of pseudo labels are selected from the temporary unlabeled data set according to a set threshold as credible pseudo labels to form a pseudo label data set; the first training model is trained for a second time using the pseudo label data set, and the loss function used in the second training includes reconstruction loss and semi-supervised classification loss based on the frequency difference between the pseudo label and the prediction, to obtain a second training model;

[0121] Fine-tuning the second training model using the labeled data set, wherein the loss function used in the fine-tuning includes a reconstruction loss and a supervised classification loss based on the true frequency difference and the predicted frequency difference, to obtain a third training model;

[0122] The second and third training processes are repeated until the prediction results of the model reach the set prediction accuracy, completing the parameter optimization of the time-frequency fusion graph convolutional network.

[0123] The mapping relationship between the cavity length change and the frequency difference is expressed as:

[0124]

[0125] Where Δf 1 ' is the second variable frequency difference of the first-stage cavity in the sensing FPI after the physical quantity to be measured changes, Δf 2 ' is the second variable frequency difference of the secondary cavity in the sensing FPI after the measured physical quantity changes, Δf 3 ' is the total second variable frequency difference in the sensor FPI after the measured physical quantity changes, ΔL 1 Indicates the change in the length of the first-stage cavity of the sensing FPI, ΔL 2 Indicates the change in the length of the sensing FPI secondary cavity, ΔL 3 Represents the total cavity length change of the sensing FPI.

[0126] Since the hardware spatial layout of the sensor system is fixed, the optical path difference between the reference FPI and the sensor FPI is fixed, that is, the detected reflection surface R 11 and R 21 The time difference of the starting moment of the reflected pulse is fixed; when the detected physical quantity acts on the sensor FPI, the variable cavity length of the first-stage cavity in the sensor FPI is changed to L' 1 , the variable cavity length of the secondary cavity becomes L' 2 , the total cavity length of the sensing FPI becomes L'3 At this time, the variable frequency differences of the primary cavity, the secondary cavity and the total in the sensing FPI become Δf 1 ', Δf 2 ' and Δf 3 '.

[0127] According to the known Δf, Δf 1 , Δf 2 , Δf 3 The mapping relationship between cavity length and frequency difference is predicted by TF-GCN to obtain the cavity length change of the sensing FPI, and then the cavity length L of each level of the sensing FPI is realized. 1 , L 2 And the total cavity length L 3 The physical quantity detection result is obtained based on the linear correspondence between the cavity length and the physical quantity to be measured, and the demodulation is completed.

[0128] A chirped pulse detection and graph convolutional neural network FPI sensing method, the method is implemented by the chirped pulse detection and graph convolutional neural network FPI sensing system, such as Figure 8 As shown, the following steps are included:

[0129] S1: Keep the physical quantity to be measured of the reference FPI unchanged, change the physical quantity to be measured of the sensing FPI multiple times, and obtain multiple interference signals containing cavity length information and frequency information through the chirp generation module and the physical quantity detection module;

[0130] The physical quantities to be measured include temperature and pressure;

[0131] Keeping the measured physical quantity of the reference FPI unchanged, after changing the measured physical quantity of the sensing FPI, the variable frequency difference of the first-stage cavity in the sensing FPI becomes Δf 1 ', the variable frequency difference of the secondary cavity becomes Δf 2 ';

[0132] The interference signal after the physical quantity to be measured is changed is obtained through the chirp generation module and the physical quantity detection module, and the physical quantity to be measured of the sensing FPI is changed multiple times to obtain multiple interference signals.

[0133] S2: pre-process all interference signals through the signal acquisition and processing module and construct the original time-frequency signal graph;

[0134] S3: The original time-frequency signal graph is predicted through the signal demodulation module based on the time-frequency fusion graph convolutional network to obtain the frequency difference after the sensor FPI physical quantity detection, and the sensor FPI cavity length change is obtained by demodulation based on the mapping relationship between the sensor FPI cavity length change and the frequency difference, and then the physical quantity detection result is determined based on the sensor FPI cavity length change.

[0135] In the mapping relationship, except for the change in the cavity length of the sensing FPI ΔL 1 , ΔL 2 and ΔL 3 All other quantities are known, so the change in the cavity length of the sensor FPI after the measured physical quantity changes can be demodulated, and then the cavity length L of the sensor FPI before the measured physical quantity changes can be obtained. 1 , L 2 and L 3 The cavity length L' of the sensing FPI after obtaining the change of the measured physical quantity 1 , L' 2 and L' 3 , and finally the physical quantity detection result is determined according to the linear correspondence between the cavity length and the physical quantity to be measured.

[0136] The FPI sensing system based on chirped pulse detection and graph convolutional neural network proposed in the present invention adopts chirped pulses as detection light, avoiding the problem that traditional demodulation methods need to use a swept frequency light source or a spectrometer, and can greatly reduce the cost of the sensing system; at the same time, the present invention adopts an FPI with known parameters as a reference, and analyzes the mapping relationship between the optical path difference and the frequency difference between the detection light and the reflected light by referring to the cavity length of the FPI, thereby completing the demodulation of the sensing FPI cavity length, and the demodulation is simple, the input parameters are few, and the complexity is low.

[0137] The present invention utilizes a graph autoencoder based on a multi-head attention mechanism to reconstruct the time-frequency graph, improves the model's ability to extract features from graph data, and enhances its ability to process time-frequency signals; and performs semi-supervised training and parameter optimization on the time-frequency fusion graph convolutional network model, which not only reduces the dependence on labeled data and improves sensing efficiency, but also enhances the model's generalization ability, making it suitable for a variety of application scenarios.

[0138] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the embodiments of the present invention is not limited thereby. Any modification, equivalent substitution and improvement made by those skilled in the art without departing from the scope and essence of the embodiments of the present invention shall be within the scope of the rights of the embodiments of the present invention.

Claims

1. A chirped pulse detection and graph convolutional neural network FPI sensing system, characterized in that: include: A chirp generation module, which is used to generate a chirped pulse incident light and split it into a first chirped pulse and a second chirped pulse; A physical quantity detection module, which is used to input the first chirped pulse and the second chirped pulse into the reference FPI and the sensing FPI respectively for physical quantity detection, to obtain the first interference signal and the second interference signal and couple them into a third interference signal; A signal acquisition and processing module, which is used to acquire the third interference signal and construct an original time-frequency signal diagram after preprocessing; The signal demodulation module is used to predict the preprocessed original time-frequency signal graph based on the time-frequency fusion graph convolutional network, obtain the frequency difference after the sensing FPI physical quantity detection, and demodulate based on the mapping relationship between the cavity length change and the frequency difference to obtain the sensing FPI cavity length change, and then determine the physical quantity detection result based on the sensing FPI cavity length change.

2. The FPI sensing system based on chirped pulse detection and graph convolutional neural network according to claim 1, characterized in that: The physical quantity detection module includes a first circulator, a second circulator, a reference FPI, a sensing FPI and an optical coupling unit; The first chirped pulse is input into the reference FPI after passing through the first circulator, and the first interference signal formed after being reflected by the reference FPI is input into the optical coupling unit through the first circulator. The second chirped pulse is input into the sensing FPI after passing through the second circulator, and the second interference signal formed after being reflected by the sensing FPI is input into the optical coupling unit through the second circulator. The optical coupling unit couples the two input interference signals and outputs the third interference signal to the signal acquisition and processing module. The reference FPI adopts a solid cavity structure with calibrated parameters; the sensing FPI includes at least one cavity with variable length, and the length of the cavity changes with the change of the physical quantity to be measured.

3. The FPI sensing system based on chirped pulse detection and graph convolutional neural network according to claim 1, characterized in that: In the signal acquisition and processing module, the third interference signal is acquired and preprocessed to construct an original time-frequency signal graph, including: After the collected third interference signal is photoelectrically converted, it is filtered and denoised, and then the interference signal after denoising is sliced, and the obtained slice vector is used as the node, the slice mode is used as the edge, and the strength of the correlation between the nodes is used as the edge weight to construct the original time-frequency signal graph of the interference signal, where the slicing mode includes pixel overlap and time window.

4. The FPI sensing system based on chirped pulse detection and graph convolutional neural network according to claim 1, characterized in that: The time-frequency fusion graph convolutional network includes a graph autoencoder, a time-frequency feature extraction module, and a frequency difference prediction module; the graph autoencoder module is used to reconstruct the graph based on the input original time-frequency signal graph to obtain a reconstructed time-frequency signal graph; the time-frequency feature extraction module is used to extract time-frequency features from the reconstructed time-frequency signal graph; the frequency domain difference prediction module is used to predict the frequency difference after each cavity changes in the sensor FPI based on the time-frequency features.

5. The FPI sensing system based on chirped pulse detection and graph convolutional neural network according to claim 4 is characterized in that: The graph autoencoder includes a graph encoder, a graph decoder, and a multi-head attention module. The slice vector sequence is used as the input of the multi-head attention module to calculate the attention weight, and the attention weight of each slice vector is introduced into the graph encoder and the graph decoder. The input original time-frequency signal graph encodes each node in the graph encoder based on the introduced attention weight to obtain the encoded feature. The encoded feature of each node is decoded in the graph decoder based on the introduced attention weight to reconstruct the graph and obtain the reconstructed time-frequency signal graph.

6. The FPI sensing system based on chirped pulse detection and graph convolutional neural network according to claim 4, characterized in that: The time-frequency feature extraction module includes at least one convolutional neural network and at least two graph convolutional networks connected in sequence, wherein the convolutional neural network is used to extract the time series features in the reconstructed time-frequency signal graph, and the graph convolutional network is used to extract the time-frequency features of each node in the reconstructed time-frequency signal graph on the basis of extracting the time series features; The frequency difference prediction module uses a fully connected layer to predict the frequency difference at different times based on the time-frequency characteristics of each node.

7. The FPI sensing system based on chirped pulse detection and graph convolutional neural network according to claim 4, characterized in that: The time-frequency fusion graph convolutional network is trained to optimize parameters before being applied. The loss function used in the training process includes the reconstruction loss between the original time-frequency signal graph and the reconstructed time-frequency signal graph, and also includes a semi-supervised classification loss constructed based on the predicted frequency difference and the actual frequency difference.

8. The FPI sensing system based on chirped pulse detection and graph convolutional neural network according to claim 7, characterized in that: The training process of the time-frequency fusion graph convolutional network includes: Prepare a dataset of original time-frequency signal graphs, and label a small part of the original time-frequency signal graphs with the real frequency difference to form a labeled dataset, and the remaining large part of the original time-frequency signal graphs constitute an unlabeled dataset; The time-frequency fusion graph convolutional network is trained for the first time using a labeled data set. The loss function used in the first training includes a reconstruction loss and a supervised classification loss based on the real frequency difference and the predicted frequency difference, thereby obtaining a first training model. The first training model is used to predict the unlabeled data set, and the prediction results are used as pseudo labels to obtain a temporary unlabeled data set with pseudo labels. A part of pseudo labels are selected from the temporary unlabeled data set according to a set threshold as credible pseudo labels to form a pseudo label data set; the first training model is trained for a second time using the pseudo label data set, and the loss function used in the second training includes reconstruction loss and semi-supervised classification loss based on the frequency difference between the pseudo label and the prediction, to obtain a second training model; Fine-tuning the second training model using the labeled data set, wherein the loss function used in the fine-tuning includes a reconstruction loss and a supervised classification loss based on the true frequency difference and the predicted frequency difference, to obtain a third training model; The second and third training processes are repeated until the prediction results of the model reach the set prediction accuracy, completing the parameter optimization of the time-frequency fusion graph convolutional network.

9. The FPI sensing system based on chirped pulse detection and graph convolutional neural network according to claim 1, characterized in that: The mapping relationship between the cavity length change and the frequency difference is expressed as: Where Δf is the fixed frequency difference of the interference signal of the reference FPI, L is the fixed cavity length of the reference FPI; i=1,2,... represents the i-th cavity in the sensing FPI, Δf i is the first variable frequency difference of the i-th cavity in the sensing FPI before the measured physical quantity changes, Δf i ' is the second variable frequency difference of the i-th cavity in the sensing FPI after the measured physical quantity changes, ΔL i It represents the change in the length of the i-th cavity of the sensing FPI before and after the physical quantity to be measured changes.

10. A FPI sensing method based on chirped pulse detection and graph convolutional neural network, characterized in that: The method adopts the graph convolutional neural network FPI sensing system according to any one of claims 1 to 9, and the method comprises the following steps: Step 1: Keep the physical quantity to be measured of the reference FPI unchanged, change the physical quantity to be measured of the sensing FPI multiple times, and obtain multiple interference signals containing cavity length information and frequency information through the chirp generation module and the physical quantity detection module; Step 2: Preprocess all interference signals through the signal acquisition and processing module and construct the original time-frequency signal graph; Step 3: The signal demodulation module predicts the original time-frequency signal graph based on the time-frequency fusion graph convolutional network to obtain the frequency difference after the sensor FPI physical quantity detection, and demodulates the sensor FPI cavity length change based on the mapping relationship between the sensor FPI cavity length change and the frequency difference to obtain the sensor FPI cavity length change, and then determines the physical quantity detection result based on the sensor FPI cavity length change.

Images