On-site manufacturing and road risk monitoring and early warning method of distributed optical fiber road sensor

Abstract

The invention belongs to the technical field of intelligent traffic, and provides a distributed optical fiber road sensor on-site manufacturing and road risk monitoring and early warning method. Comprising the steps of optical fiber unit manufacturing, concave shallow slot cutting, sensor construction, data acquisition, road surface state data conversion, scattered light time domain signal interval division, frequency domain conversion and filtering processing, multi-mode feature extraction and road surface condition recognition. According to the invention, the concave shallow slot is cut at the edge of the road, and the flat ribbon-shaped optical fiber unit is fixed in the shallow slot by using polyurea, so that the multifunctional distributed optical fiber road sensor which is integrated with the road surface, high in sensitivity, small in size, simple to install, convenient and rapid to maintain and repair and low in cost is constructed; through a zero crossing point and a noise threshold value, normal time domain signal interval judgment is optimized, and the effect of subsequent filtering processing is improved; a road surface condition identification data model is constructed through a multi-modal fusion technology, and the identification precision of multiple road surface abnormal conditions and the positioning precision of abnormal events are improved.

Description

Technical Field

[0001] The present invention relates to the field of smart transportation technology, and in particular to an on-site production method of a distributed optical fiber road sensor and a road risk monitoring and early warning method. Background Art

[0002] With the dramatic increase in the number of urban roads and highway networks, many roads reaching the end of their lifespans or even exceeding their lifespans, and the increasing number of traffic accidents caused by natural geological disasters and man-made construction damage, road safety monitoring has become increasingly important. Traditional road safety relies primarily on manual patrols, which suffer from limited monitoring range, poor real-time performance, ineffectiveness, and high costs. Distributed fiber optic sensing technology, with its advantages of high sensitivity, passive sensors, inherent safety, wide monitoring range, strong real-time performance, and resistance to electromagnetic interference, has gradually become a research hotspot in the field of road risk monitoring and early warning.

[0003] However, existing communication optical cables have problems such as large size, easy damage, complex installation, and high cost of pre-buried construction or road cutting during installation and use, which limits their application in actual projects. In addition, most existing distributed optical fiber signal demodulation systems use conventional scattering principles to calculate temperature, pressure, stress, etc., but lack corresponding processing methods for interference from the environment surrounding the optical fiber sensor, resulting in a lack of reliability of the demodulated data. Summary of the Invention

[0004] In order to overcome the shortcomings of the existing technology, the purpose of the present invention is to provide an on-site production method of a distributed fiber optic road sensor and a road risk monitoring and early warning method to solve the problems of traditional distributed fiber optic application limitations, poor background signal interval division effect, and poor model recognition effect on multiple road conditions.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for on-site fabrication of a distributed optical fiber road sensor and road risk monitoring and early warning, comprising:

[0007] A non-metallic flat ribbon optical fiber unit with 5 or more cores and a thickness of 1.0 mm is made using epoxy glass fiber resin; the non-metallic flat ribbon optical fiber unit includes: 1 core multimode optical fiber and 4 cores single-mode optical fibers;

[0008] Use a diamond blade to cut a shallow concave groove with a width and depth of 4 mm at a distance of 50 mm from the edge of the target road surface;

[0009] A 1mm thick layer of flexible polyurea is sprayed on the bottom of the concave shallow groove, and the non-metallic flat ribbon optical fiber unit is evenly and flatly fixed on the polyurea surface at the bottom of the concave shallow groove. Rigid polyurea with a temperature resistance of 200°C is injected into the concave shallow groove until it is flush with the road surface. After the polyurea is cured, a distributed optical fiber road sensor is obtained.

[0010] Using the distributed optical fiber road sensor to collect time and space data on the target road surface without blind spots to obtain road surface status data;

[0011] The road surface state data is divided into intervals using a zero-crossing point and a noise threshold to obtain a road surface event time domain signal and a normal time domain signal;

[0012] Performing Fourier transformation and filtering on the road surface event time domain signal according to the normal time domain signal to obtain a road surface abnormal event characteristic signal model to establish a characteristic event database;

[0013] Performing multimodal feature extraction on the feature event database to obtain multimodal fusion analysis data;

[0014] The multimodal fusion analysis data is input into a pre-trained road condition recognition model for training to obtain various road abnormality recognition results.

[0015] Preferably, a layer of flexible polyurea with a thickness of 1 mm is sprayed on the bottom of the concave shallow groove, and the non-metallic flat ribbon optical fiber unit is evenly and flatly fixed on the polyurea surface at the bottom of the concave shallow groove. Rigid polyurea with a temperature resistance of 200°C is injected into the concave shallow groove until it is flush with the road surface. After the polyurea is cured, a distributed optical fiber road sensor is obtained, comprising:

[0016] A bracket is made, two polyurea spray guns are fixed to the front and rear ends of the bracket, and a positioner is installed in front of the nozzle of the polyurea spray gun at the rear end to obtain an optical fiber laying device; the width of the positioner is 0.5 mm less than the width of the concave shallow groove;

[0017] Mounting the optical fiber ribbon reel of the flat optical fiber ribbon unit on the bracket;

[0018] During construction, impurities in the concave shallow groove are cleaned, and a layer of polyurea with a thickness of 1 mm is sprayed on the bottom of the concave shallow groove using the polyurea spray gun at the front end. The flat ribbon optical fiber unit is placed at the bottom of the concave shallow groove using the positioner, and polyurea is injected into the concave shallow groove using the polyurea spray gun at the rear end until it is flush with the road surface. After the polyurea is cured, a distributed optical fiber road sensor is obtained.

[0019] Preferably, the distributed optical fiber road sensor is used to collect data on the target road surface in time and space without blind spots to obtain road surface status data, including:

[0020] An optical time domain reflectometer and a multi-channel synchronous acquisition card are respectively connected to the head and tail of the distributed optical fiber road sensor;

[0021] Using the optical time domain reflectometer, a laser with a wavelength of 1550 nm, a pulse width of 50 ns to 100 ns, and a sampling frequency of 3 kHz is emitted into the single-mode optical fiber, and a laser with a wavelength less than 1000 nm, a pulse width greater than 100 ns, and a sampling frequency of 1.5 kHz is emitted into the multimode optical fiber;

[0022] Using the multi-channel synchronous acquisition card to synchronously demodulate the optical signals output by the single-mode optical fiber and the multi-mode optical fiber to obtain the road surface state data;

[0023] When the zero-crossing rate of the road surface state data within a unit time window is greater than a preset threshold value of normal road surface conditions, the sampling frequencies of the single-mode optical fiber and the multi-mode optical fiber are adjusted to 5 kHz and 2.5 kHz, respectively.

[0024] Preferably, the road surface state data is divided into intervals using a zero-crossing point and a noise threshold to obtain a road surface event time domain signal and a normal time domain signal, including:

[0025] Extracting an interval greater than the noise threshold from the road surface state data to obtain an initial interval;

[0026] Extracting zero-crossing points in the road surface state data to obtain a zero-crossing point set;

[0027] Determining the interval range of the two zero-crossing points closest to the initial interval as a road event time domain interval based on the zero-crossing point set, and determining the interval range of the road state data excluding the road event time domain interval as a background time domain interval;

[0028] The road surface state data is divided into intervals according to the road surface event time domain interval and the background time domain interval to obtain the road surface event time domain signal and the normal time domain signal.

[0029] Preferably, the road event time domain signal is subjected to Fourier transformation and filtering processing according to the normal time domain signal to obtain a road abnormal event characteristic signal model to establish a characteristic event database, including:

[0030] Performing frequency domain conversion on the road event time domain signal and the normal time domain signal using fast Fourier transform to obtain a road event frequency domain signal and a background frequency domain signal;

[0031] A difference calculation is performed on the road event frequency domain signal and the background frequency domain signal to obtain the characteristic event database.

[0032] Preferably, multimodal feature extraction is performed on the feature event database to obtain multimodal fusion analysis data, including:

[0033] The high-frequency energy distribution in each time window of the characteristic event database is extracted to obtain the STFT time-frequency vector; the expression of the STFT time-frequency vector is:

[0034] X stft =[E 100-200 ,E 200-300 ,E 300-500 ,Peak_freq,Energy_std];

[0035] Among them, X stft is the STFT time-frequency vector; E a-b Represents the normalized energy in the frequency band [a, b]; Peak_freq represents the main frequency of the impulse signal; Energy_std represents the standard deviation of the energy in the frequency band from 100Hz to 500Hz;

[0036] Perform wavelet decomposition on the characteristic event database and calculate the energy entropy at each scale to obtain a wavelet energy entropy vector; the expression of the wavelet energy entropy vector is: wavelet =[H1,H2,H3]; where X wavelet is the wavelet energy entropy vector; H1, H2, and H3 are the energy entropies of the 1st, 2nd, and 3rd layers of the wavelet decomposition of the characteristic event database respectively;

[0037] Extract the time domain instantaneous characteristics of the road event frequency domain signal to obtain an instantaneous feature vector; the expression of the instantaneous feature vector is: transient =[t rise ,t duration ]; where X transient is the instantaneous eigenvector; t rise The time it takes for the signal to rise from 10% to 90% of its peak value; t duration is the duration of the impact signal;

[0038] The STFT time-frequency vector, the wavelet energy entropy vector, and the instantaneous feature vector are fused to obtain a mechanical shock class vector.

[0039] Preferably, multimodal feature extraction is performed on the feature event database to obtain multimodal fusion analysis data, including:

[0040] The energy proportion and modal frequency of the low-frequency signal in the characteristic event database are calculated to obtain a low-frequency energy eigenvector; the expression of the low-frequency energy eigenvector is:

[0041] X lowfreq =[E low ,f mode1 ,f mode2 ];

[0042] Among them, X lowfreq is the low-frequency energy eigenvector; E low is the energy proportion of the frequency band less than 100 Hz; f mode1 is the first-order vibration mode frequency; f mode2 is the second-order vibration mode frequency;

[0043] Calculate the variance and skewness of the characteristic event database to obtain the strain-frequency coupling characteristic vector; the expression of the strain-frequency coupling characteristic vector is: X strain_freq =[Var_f,Skew_f]; where X strain_freq is the strain-frequency coupling eigenvector; Var_f is the variance of the signal frequency; Skew_f is the skewness of the signal frequency;

[0044] Extract the gradient change of the temperature channel of the characteristic event database of the two time periods to obtain the temperature anomaly feature vector; the expression of the temperature anomaly feature vector is: temp =[ΔT1,ΔT2]; where X temp is the temperature anomaly characteristic vector; ΔT1 and ΔT2 are the temperature change rates in the first time period and the second time period respectively; the second time period is greater than the first time period;

[0045] The low-frequency energy eigenvector, the strain-frequency coupling eigenvector, and the temperature anomaly eigenvector are fused to obtain a structural deformation vector.

[0046] Preferably, performing multimodal feature extraction on the feature event database to obtain multimodal fusion analysis data further includes:

[0047] The energy entropy of the high frequency band and the power spectrum density of the low frequency band of the characteristic event database are extracted to obtain a first environmental feature vector; the expression of the first environmental feature vector is:

[0048] X1=[H 1_high ,PSD 1_low ];

[0049] Wherein, X1 is the first environmental feature vector; H 1_high Energy entropy from 50Hz to 200Hz; PSD 1_lowis the power spectrum density value of the frequency band less than 30Hz;

[0050] The low-frequency energy proportion and high-frequency pulse frequency of the characteristic event database and the signal duration and low-frequency energy rising slope of the road event frequency domain signal are extracted to obtain a second environmental feature vector; the expression of the second environmental feature vector is:

[0051] X2=[E 2_low ,f impulse ,t duration ,Slope_E];

[0052] Wherein, X2 is the second environment feature vector; E 2_low is the energy proportion of the frequency band less than 20Hz; f impulse is the pulse frequency from 50Hz to 200Hz; Slope_E is the rate of change of energy in the frequency band less than 20Hz;

[0053] splicing the mechanical impact vector, the structural deformation vector, the first environmental feature vector, and the second environmental feature vector to obtain an auxiliary feature matrix;

[0054] Converting the characteristic event database corresponding to the multimode optical fiber and the single-mode optical fiber into a multimode optical fiber time-frequency diagram and a single-mode optical fiber time-frequency diagram;

[0055] The multimode optical fiber time-frequency diagram, the single-mode optical fiber time-frequency diagram, and the auxiliary feature matrix are fused to obtain the multimodal fusion analysis data.

[0056] Preferably, the training process of the road condition recognition model includes:

[0057] Pre-collecting the multimodal fusion analysis data and adding real road condition labels to the multimodal fusion analysis data; the real road condition labels include: collision, vehicle rollover, tire blowout, road collapse, sudden heavy rainfall, road icing, falling rocks, mudslide, fallen objects, and explosions;

[0058] Build the basic model architecture;

[0059] Using the basic model architecture, the multimode optical fiber time-frequency graph is sequentially subjected to lightweight convolution, hole convolution processing, and group convolution to obtain a multimode optical fiber characteristic graph;

[0060] The spatial attention gating mechanism of the basic model architecture is used to perform maximum pooling, average pooling, linear transformation, and Sigmoid function activation on the multimode fiber feature map to obtain an attention weight;

[0061] Multiplying the attention weight and the multimode fiber feature map and adjusting the size to a preset fixed size through an adaptive pooling operation to obtain a multimode feature;

[0062] Using the basic model architecture, lightweight convolution and time dimension pooling are performed on the single-mode optical fiber time-frequency map to obtain a single-mode optical fiber feature map;

[0063] Performing modal interaction on the single-mode optical fiber feature map and adjusting the size to the fixed size through an adaptive pooling operation to obtain a single-mode feature;

[0064] Using the temporal convolutional network of the basic model architecture to perform causal convolution processing on the auxiliary feature matrix and perform dimensionality reduction through a fully connected layer to obtain a time series vector;

[0065] The basic model architecture is used to perform dynamic scaling gating processing, feature splicing and dimensionality reduction processing on the multi-mode features, the single-mode features and the time series vector respectively to obtain fused features.

[0066] Preferably, the training process of the road condition recognition model further includes:

[0067] Utilizing the Softmax activation function to calculate the road condition probability distribution of the fused features, and obtaining a network prediction result;

[0068] The loss function is calculated using the network prediction result and the real road condition label to obtain the model loss value; the expression of the loss function is:

[0069]

[0070] Among them, L cls is the model loss value; M is the number of batch training samples; N is the total number of road condition categories; y i,c is the true label of the i-th sample, which is 1 if it belongs to category c, otherwise it is 0; is the predicted probability that the i-th sample belongs to category c in the network prediction result;

[0071] The basic model architecture is parameter optimized and iterated using the model loss value according to the minimum gradient descent strategy to obtain the trained road condition recognition model.

[0072] The present invention discloses the following technical effects:

[0073] The present invention provides a method for on-site production of distributed fiber optic road sensors and road risk monitoring and early warning. By cutting a concave shallow groove on the edge of the road and fixing a flat ribbon fiber unit in the groove with polyurea, the problems of traditional distributed fiber optic sensors such as large size, easy damage, complex installation, and high cost of pre-buried construction or road cutting are solved, and the production of distributed fiber optic road sensors with small size, simple installation, and low cost is achieved. The zero-crossing point and noise threshold are used to solve the problem of poor background signal interval division effect, and the interval optimization of the preliminary division result using the noise threshold is achieved by using the zero-crossing point. The road condition recognition data model is constructed by multimodal fusion technology, which solves the problem of poor model recognition effect for multiple road conditions and achieves simultaneous and accurate recognition of vehicle collisions, vehicle rollovers, tire blowouts, road collapses, sudden heavy rainfall, road icing, falling rocks, mudslides, fallen objects, and object explosions. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0075] Figure 1 A schematic diagram of the on-site production of a distributed optical fiber road sensor and the road risk monitoring and early warning process provided by an embodiment of the present invention;

[0076] Figure 2 A schematic diagram of the manufacturing process of a distributed optical fiber road sensor provided in an embodiment of the present invention;

[0077] Figure 3 A schematic diagram of a road surface status data collection process according to an embodiment of the present invention;

[0078] Figure 4 A schematic diagram of the interval division process provided by an embodiment of the present invention;

[0079] Figure 5 A schematic diagram of the frequency domain conversion and filtering process provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0080] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0081] The purpose of the present invention is to provide a method for on-site production of distributed optical fiber road sensors and road risk monitoring and early warning, so as to solve the problems of traditional distributed optical fiber application limitations, poor background signal interval division effect, and poor model recognition effect on multiple road conditions.

[0082] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0083] Figure 1 A schematic diagram of the on-site production of distributed optical fiber road sensors and road risk monitoring and early warning process provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a method for on-site production of a distributed optical fiber road sensor and road risk monitoring and early warning, comprising:

[0084] Step 100: Using epoxy glass fiber resin, a non-metallic flat ribbon optical fiber unit having 5 or more cores and a thickness of 1.0 mm is manufactured; the non-metallic flat ribbon optical fiber unit includes: 1 core multimode optical fiber and 4 cores single-mode optical fibers;

[0085] Step 200: Using a diamond blade, a shallow concave groove with a width and depth of 4 mm is cut on the target road surface at a distance of 50 mm from the edge;

[0086] Step 300: spraying a 1mm thick layer of flexible polyurea onto the bottom of the concave shallow groove, and evenly and flatly fixing the non-metallic flat ribbon optical fiber unit onto the polyurea surface at the bottom of the concave shallow groove. Rigid polyurea with a temperature resistance of 200°C is then injected into the concave shallow groove until it is flush with the road surface. After the polyurea is cured, a distributed optical fiber road sensor is obtained.

[0087] Step 400: Using the distributed optical fiber road sensor to collect time-space data on the target road surface without blind spots, to obtain road surface status data;

[0088] Step 500: dividing the road surface state data into intervals using a zero-crossing point and a noise threshold to obtain a road surface event time domain signal and a normal time domain signal;

[0089] Step 600: Performing Fourier transformation and filtering on the road event time domain signal according to the normal time domain signal to obtain a road abnormal event characteristic signal model to establish a characteristic event database;

[0090] Step 700: extracting multimodal features from the feature event database to obtain multimodal fusion analysis data;

[0091] Step 800: Input the multimodal fusion analysis data into a pre-trained road condition recognition model for training to obtain various road abnormality recognition results.

[0092] refer to Figure 2 A 1mm thick layer of flexible polyurea is sprayed onto the bottom of the concave shallow groove, and the non-metallic flat ribbon optical fiber unit is evenly and flatly fixed on the polyurea surface at the bottom of the concave shallow groove. Rigid polyurea with a temperature resistance of 200°C is injected into the concave shallow groove until it is flush with the road surface. After the polyurea is cured, a distributed optical fiber road sensor is obtained, comprising:

[0093] Step 301: Prepare a bracket, fix two polyurea spray guns at the front and rear ends of the bracket, and install a positioner in front of the nozzle of the polyurea spray gun at the rear end to obtain an optical fiber laying device; the width of the positioner is 0.5 mm less than the width of the concave shallow groove;

[0094] Step 302: Installing the optical fiber ribbon reel of the flat optical fiber ribbon unit on the bracket;

[0095] Step 303: During construction, impurities in the concave shallow groove are cleaned, and a layer of polyurea with a thickness of 1 mm is sprayed on the bottom of the concave shallow groove using the polyurea spray gun at the front end. The flat ribbon optical fiber unit is placed at the bottom of the concave shallow groove using the positioner, and polyurea is injected into the concave shallow groove using the polyurea spray gun at the rear end until it is flush with the road surface. After the polyurea is cured, a distributed optical fiber road sensor is obtained.

[0096] refer to Figure 3 , using the distributed optical fiber road sensor to collect time and space data on the target road surface without blind spots to obtain road surface status data, including:

[0097] Step 401: Connect an optical time domain reflectometer and a multi-channel synchronous acquisition card to the head and tail of the distributed optical fiber road sensor respectively;

[0098] Step 402: Using the optical time domain reflectometer, emit a laser with a wavelength of 1550 nm, a pulse width between 50 ns and 100 ns, and a sampling frequency of 3 kHz into the single-mode optical fiber, and emit a laser with a wavelength less than 1000 nm, a pulse width greater than 100 ns, and a sampling frequency of 1.5 kHz into the multimode optical fiber;

[0099] Step 403: synchronously demodulating the optical signals output by the single-mode optical fiber and the multi-mode optical fiber using the multi-channel synchronous acquisition card to obtain the road surface state data;

[0100] Step 404: When the zero-crossing rate of the road surface state data within a unit time window is greater than a preset threshold value for normal road conditions, the sampling frequencies of the single-mode optical fiber and the multi-mode optical fiber are adjusted to 5 kHz and 2.5 kHz, respectively.

[0101] refer to Figure 4 , using the zero-crossing point and the noise threshold to divide the road surface state data into intervals, and obtaining a road surface event time domain signal and a normal time domain signal, including:

[0102] Step 501: extracting intervals greater than the noise threshold in the road surface state data to obtain an initial interval;

[0103] Step 502: extracting zero-crossing points in the road surface state data to obtain a zero-crossing point set;

[0104] Step 503: Determine the interval range of the two zero-crossing points closest to the initial interval as the road event time domain interval based on the zero-crossing point set, and determine the interval range of the road state data excluding the road event time domain interval as the background time domain interval;

[0105] Step 504: Divide the road surface state data into intervals according to the road surface event time domain interval and the background time domain interval to obtain the road surface event time domain signal and the normal time domain signal.

[0106] refer to Figure 5 , performing Fourier transform and filtering processing on the road event time domain signal according to the normal time domain signal to obtain a road abnormal event characteristic signal model to establish a characteristic event database, including:

[0107] Step 601: Perform frequency domain conversion on the road event time domain signal and the normal time domain signal using fast Fourier transform to obtain a road event frequency domain signal and a background frequency domain signal;

[0108] Step 602: performing a difference calculation on the road event frequency domain signal and the background frequency domain signal to obtain the feature event database.

[0109] Specifically, multimodal feature extraction is performed on the feature event database to obtain multimodal fusion analysis data, including:

[0110] The high-frequency energy distribution in each time window of the characteristic event database is extracted to obtain the STFT time-frequency vector; the expression of the STFT time-frequency vector is:

[0111] X stft =[E 100-200 ,E 200-300 ,E 300-500 ,Peak_freq,Energy_std];

[0112] Among them, X stft is the STFT time-frequency vector; E a-b Represents the normalized energy in the frequency band [a, b]; Peak_freq represents the main frequency of the impulse signal; Energy_std represents the standard deviation of the energy in the frequency band from 100Hz to 500Hz;

[0113] Perform wavelet decomposition on the characteristic event database and calculate the energy entropy at each scale to obtain a wavelet energy entropy vector; the expression of the wavelet energy entropy vector is: wavelet =[H1,H2,H3]; where X wavelet is the wavelet energy entropy vector; H1, H2, and H3 are the energy entropies of the 1st, 2nd, and 3rd layers of the wavelet decomposition of the characteristic event database respectively;

[0114] Extract the time domain instantaneous characteristics of the road event frequency domain signal to obtain an instantaneous feature vector; the expression of the instantaneous feature vector is: transient =[t rise ,t duration ]; where X transient is the instantaneous eigenvector; t rise The time it takes for the signal to rise from 10% to 90% of its peak value; t duration is the duration of the impact signal;

[0115] The STFT time-frequency vector, the wavelet energy entropy vector, and the instantaneous feature vector are fused to obtain a mechanical shock class vector.

[0116] Furthermore, multimodal feature extraction is performed on the feature event database to obtain multimodal fusion analysis data, including:

[0117] The energy proportion and modal frequency of the low-frequency signal in the characteristic event database are calculated to obtain a low-frequency energy eigenvector; the expression of the low-frequency energy eigenvector is:

[0118] X lowfreq =[E low ,f mode1 ,f mode2 ];

[0119] Among them, X lowfreq is the low-frequency energy eigenvector; E low is the energy proportion of the frequency band less than 100 Hz; f mode1 is the first-order vibration mode frequency; f mode2 is the second-order vibration mode frequency;

[0120] Calculate the variance and skewness of the characteristic event database to obtain the strain-frequency coupling characteristic vector; the expression of the strain-frequency coupling characteristic vector is: X strain_freq =[Var_f,Skew_f]; where X strain_freq is the strain-frequency coupling eigenvector; Var_f is the variance of the signal frequency; Skew_f is the skewness of the signal frequency;

[0121] Extract the gradient change of the temperature channel of the characteristic event database of the two time periods to obtain the temperature anomaly feature vector; the expression of the temperature anomaly feature vector is: temp =[ΔT1,ΔT2]; where X temp is the temperature anomaly characteristic vector; ΔT1 and ΔT2 are the temperature change rates in the first time period and the second time period respectively; the second time period is greater than the first time period;

[0122] The low-frequency energy eigenvector, the strain-frequency coupling eigenvector, and the temperature anomaly eigenvector are fused to obtain a structural deformation vector.

[0123] Furthermore, performing multimodal feature extraction on the feature event database to obtain multimodal fusion analysis data also includes:

[0124] The energy entropy of the high frequency band and the power spectrum density of the low frequency band of the characteristic event database are extracted to obtain a first environmental feature vector; the expression of the first environmental feature vector is:

[0125] X1=[H 1_high ,PSD 1_low ];

[0126] Wherein, X1 is the first environmental feature vector; H 1_high Energy entropy from 50Hz to 200Hz; PSD 1_low is the power spectrum density value of the frequency band less than 30Hz;

[0127] The low-frequency energy proportion and high-frequency pulse frequency of the characteristic event database and the signal duration and low-frequency energy rising slope of the road event frequency domain signal are extracted to obtain a second environmental feature vector; the expression of the second environmental feature vector is:

[0128] X2=[E 2_low ,f impulse ,t duration ,Slope_E];

[0129] Wherein, X2 is the second environment feature vector; E 2_low is the energy proportion of the frequency band less than 20Hz; f impulseis the pulse frequency from 50Hz to 200Hz; Slope_E is the rate of change of energy in the frequency band less than 20Hz;

[0130] splicing the mechanical impact vector, the structural deformation vector, the first environmental feature vector, and the second environmental feature vector to obtain an auxiliary feature matrix;

[0131] Converting the characteristic event database corresponding to the multimode optical fiber and the single-mode optical fiber into a multimode optical fiber time-frequency diagram and a single-mode optical fiber time-frequency diagram;

[0132] The multimode optical fiber time-frequency diagram, the single-mode optical fiber time-frequency diagram, and the auxiliary feature matrix are fused to obtain the multimodal fusion analysis data.

[0133] Specifically, the training process of the road condition recognition model includes:

[0134] Pre-collecting the multimodal fusion analysis data and adding real road condition labels to the multimodal fusion analysis data; the real road condition labels include: collision, vehicle rollover, tire blowout, road collapse, sudden heavy rainfall, road icing, falling rocks, mudslide, fallen objects, and explosions;

[0135] Build the basic model architecture;

[0136] Using the basic model architecture, the multimode optical fiber time-frequency graph is sequentially subjected to lightweight convolution, hole convolution processing, and group convolution to obtain a multimode optical fiber characteristic graph;

[0137] The spatial attention gating mechanism of the basic model architecture is used to perform maximum pooling, average pooling, linear transformation, and Sigmoid function activation on the multimode fiber feature map to obtain an attention weight;

[0138] Multiplying the attention weight and the multimode fiber feature map and adjusting the size to a preset fixed size through an adaptive pooling operation to obtain a multimode feature;

[0139] Using the basic model architecture, lightweight convolution and time dimension pooling are performed on the single-mode optical fiber time-frequency map to obtain a single-mode optical fiber feature map;

[0140] Performing modal interaction on the single-mode optical fiber feature map and adjusting the size to the fixed size through an adaptive pooling operation to obtain a single-mode feature;

[0141] Using the temporal convolutional network of the basic model architecture to perform causal convolution processing on the auxiliary feature matrix and perform dimensionality reduction through a fully connected layer to obtain a time series vector;

[0142] The basic model architecture is used to perform dynamic scaling gating processing, feature splicing and dimensionality reduction processing on the multi-mode features, the single-mode features and the time series vector respectively to obtain fused features.

[0143] Furthermore, the training process of the road condition recognition model further includes:

[0144] Utilizing the Softmax activation function to calculate the road condition probability distribution of the fused features, and obtaining a network prediction result;

[0145] The loss function is calculated using the network prediction result and the real road condition label to obtain the model loss value; the expression of the loss function is:

[0146]

[0147] Among them, L cls is the model loss value; M is the number of batch training samples; N is the total number of road condition categories; y i,c is the true label of the i-th sample, which is 1 if it belongs to category c, otherwise it is 0; is the predicted probability that the i-th sample belongs to category c in the network prediction result;

[0148] The basic model architecture is parameter optimized and iterated using the model loss value according to the minimum gradient descent strategy to obtain the trained road condition recognition model.

[0149] Specifically, the fiber ribbon unit is prepared by using epoxy glass fiber resin to create a flat fiber ribbon unit with at least five cores. The fiber ribbon unit is installed by using a diamond blade to cut a shallow, concave groove 4mm wide and 4mm deep in the concrete or asphalt pavement 50mm from the outside of the road (close to the edge of the road to avoid being crushed by wheels). The groove is then cleaned with a hair dryer.

[0150] Furthermore, a bracket is constructed and two polyurea spray guns are secured to the bracket, one in front of the other. A locator is installed in front of the nozzle of the rear polyurea spray gun to position the fiber ribbon unit. The width of the locator is slightly smaller than the width of the road surface groove. During construction, the locator guides the cart along the groove. A positioning groove, slightly larger than the width of the fiber ribbon unit, is provided in the center of the bottom of the locator. The locator ensures that the fiber ribbon unit is centered in the road surface groove. A shaft is mounted on the bracket to accommodate the fiber ribbon reel, facilitating placement of the fiber ribbon unit. The bracket and spray guns are mounted on a cart for easy operation.

[0151] Furthermore, during construction, the polyurea spray gun in front is first used to spray the bottom layer of polyurea at the bottom of the groove, and the optical fiber ribbon unit is simultaneously placed in the groove. The optical fiber ribbon unit is fixed by the viscosity of the bottom layer of polyurea sprayed by the polyurea gun in front. After the optical fiber ribbon unit enters the groove, the polyurea spray gun behind the positioner simultaneously injects polyurea into the groove to fix the optical fiber unit, and adjusts the amount of polyurea to ensure that the polyurea is level with the road surface after curing, thereby completing the encapsulation of the optical fiber unit.

[0152] Specifically, without damaging or re-casting the existing pavement, fiber optic sensors are encapsulated in shallow trenches in the road surface. The optical fibers are connected to distributed fiber optic temperature, vibration, and strain signal demodulators and data processing terminals, enabling the following functions: Real-time temperature monitoring along the entire road surface to prevent icing; vibration monitoring to prevent damage to roads or underground pipelines caused by excavation, pipe jacking, or shield tunneling; and early warning of road surface anomalies caused by subsidence, heave deformation, landslides, and displacement, reducing traffic accidents.

[0153] Preferably, a shallow groove of 4*4mm is opened on the road surface without affecting the quality, performance and life of the road; epoxy glass fiber resin and polyurea are used for encapsulation to improve the mechanical strength and durability of the ribbon optical fiber unit, such as waterproof, moisture-proof, and anti-ultraviolet aging; the optical fiber sensing unit is encapsulated in the groove of the road surface and integrated with the road surface, so that the force response and vibration transmission effects are faster and more sensitive; the grooves opened on the asphalt or cement road are directly used to protect the optical fiber sensing unit, which not only saves the packaging materials of the optical fiber sensor, but also greatly reduces the volume of the distributed optical fiber sensor; the installation process is simplified by grooving and polyurea fixation, which reduces the construction difficulty and facilitates the maintenance and replacement of the optical fiber unit; the optical fiber sensor of this embodiment can monitor temperature, vibration and strain at the same time, fully covering the road risk monitoring needs, and through distributed optical fiber sensing technology, it can realize real-time monitoring of the entire road line and improve the early warning capability. Table 1 is the specific data of some parameters provided in this embodiment.

[0154] Table 1

[0155]

[0156] Specifically, a high-precision OTDR (Optical Time Domain Reflectometer) module is used, which supports simultaneous emission of multiple wavelengths. The single-mode channel captures dynamic events such as vehicle passage in real time, while the multi-mode channel balances response speed and signal-to-noise ratio. By integrating a multi-channel synchronous acquisition card, the signals of the four-core single-mode fiber and the one-core multi-mode fiber are synchronously demodulated to avoid missed events caused by time misalignment. The initial sensing fiber data collected by the data acquisition card is the initial time domain signal. The fiber data collected by the data acquisition card is converted into road surface status data corresponding to different locations on the monitored road section through serial-to-parallel conversion, as follows:

[0157] The road surface condition data generated by a single light pulse at different spatial positions of the sensing optical fiber is collected by a data acquisition card;

[0158] The Rayleigh scattered light signals collected at different spatial positions are converted serially to parallel to obtain the road surface state data at different moments at the determined spatial positions;

[0159] The road surface condition data at different times at the determined spatial position are subjected to sliding window truncation processing using preset sliding window parameters to obtain a set of road surface condition data at a determined spatial position and time.

[0160] Furthermore, a double screening mechanism is used to accurately locate the event interval:

[0161] Noise threshold preliminary screening: By extracting the interval greater than the noise threshold in the signal, the initial signal segment containing potential road events is quickly locked in, excluding obvious background signals;

[0162] Zero-crossing points accurately define boundaries: Signal zero-crossing points are further extracted (i.e., intersections where the signal changes from positive to negative or negative to positive, determined by multiplying the preceding and following points). Using this set of zero-crossing points, the two closest to the initial interval are identified as event boundaries. This is because road events (such as vehicle impacts and structural vibrations) often cause significant fluctuations in signal amplitude. Changes in zero-crossing points accurately reflect the start and end of an event, preventing misjudgment of intervals due to noise fluctuations.

[0163] Preferably, signal segments excluding the event interval are defined as normal time-domain signals, effectively separating normalized environmental signals from sudden road event signals. The threshold used in this process is dynamically calculated from the signal data within the sliding time window and is not a fixed constant. This avoids degradation of effectiveness in multiple scenarios. A quantile method is used, using the 95th percentile of the signal amplitude within the window as the threshold to ensure coverage of the normal noise fluctuation range. Furthermore, when the system detects significant changes in signal characteristics such as zero-crossing rate and energy distribution within multiple consecutive time windows, it automatically triggers threshold recalculation, avoiding lags in fixed thresholds when the environment suddenly changes.

[0164] Furthermore, the frequency domain information containing the external vibration event is filtered according to the background spectrum to determine the frequency domain information of the collected signal within its characteristic frequency band; wherein, the background noise spectrum is obtained by performing frequency domain analysis on the frequency domain information when no external event occurs, that is, the frequency band corresponding to the above-mentioned background signal; the frequency domain information containing the external event is subtracted from the frequency spectrum obtained when no event occurred before the corresponding position, to obtain the filtered frequency domain information.

[0165] Preferably, the traditional single mode is difficult to fully describe the multi-dimensional characteristics of complex road events (such as car crashes, road collapses, icing, etc.). Multimodal fusion can avoid the problem of missed or misjudgment of a single mode by integrating multi-dimensional data, covering multiple types of features such as mechanical impact, structural deformation, and environmental changes. The STFT time-frequency vector is used to capture the high-frequency energy distribution and frequency concentration characteristics of short-term impact events such as vehicle collisions and tire blowouts; the wavelet energy entropy vector reflects the time-frequency localization characteristics of the signal and is sensitive to the complexity of the time-frequency distribution of the impact signal; the instantaneous eigenvector can quantify the transient characteristics of the impact signal; the combination of the above three vectors can fully cover the frequency band characteristics of mechanical impact events. For structural deformation events, it is necessary to integrate low-frequency energy eigenvectors, strain-frequency coupling eigenvectors, and temperature anomaly eigenvectors. For environmental events, the main consideration is the impact of weather.

[0166] Specifically, in order to adapt to the input of multimodal fusion data, this embodiment designs a three-branch network architecture, which performs independent feature analysis on multimode fiber data, single-mode fiber data, and the auxiliary feature matrix data composed of the aforementioned various vectors, and then performs fusion analysis. For multimode fiber data: a 5×5 large convolution kernel plus a hole convolution is used to capture the time-frequency distribution of the broadband impact signal; for single-mode fiber data: a 3×3 convolution kernel + time dimension pooling is used to enhance the low-frequency structural deformation characteristics; the auxiliary feature matrix data is used as auxiliary identification data and is processed using a time convolution network (TCN). The specific instructions are as follows:

[0167] 1) Multi-mode feature branch

[0168] The multimode fiber time-frequency image is first processed using GhostConv (a lightweight convolution block) and dilated convolution. GhostConv reduces the number of parameters while maintaining feature expression capabilities, making it suitable for the computing power constraints of edge computing nodes. Dilated convolution uses a large 5×5 convolution kernel and a dilation ratio of 2 to capture the time-frequency distribution of broadband impulse signals, expand the receptive field, and enhance the perception of high-frequency signals.

[0169] Then, group convolution is performed to group the input channels for convolution operation, which further reduces the amount of calculation and can learn the feature representations within different groups;

[0170] A spatial attention gating mechanism is applied to the feature map. By calculating the maximum pooling and average pooling results of the feature map, applying a linear transformation and a sigmoid activation function, the attention weight is obtained, and then multiplied by the feature map. This step can focus on the effective signal area in the fiber optic laying area, suppress road edge noise, and highlight important feature information.

[0171] Finally, an adaptive pooling operation is performed to adjust the size of the feature map to a fixed size for subsequent feature fusion.

[0172] 2) Single-mode feature branch

[0173] The single-mode fiber time-frequency graph is processed using GhostConv and time-dimension pooling. GhostConv also reduces the number of parameters, while time-dimension pooling uses a pooling kernel to enhance the temporal information of low-frequency structural deformation features and capture the temporal correlation of strain signals.

[0174] By performing modal interaction through methods such as dot multiplication, the features extracted by different convolutional layers are fused to mine the feature correlation within the single-mode fiber signal;

[0175] Pool the feature map to a fixed size in preparation for subsequent fusion.

[0176] 3) Timing feature branch

[0177] The auxiliary feature matrix is input into the temporal convolutional network (TCN). TCN uses causal convolution. By setting different dilation values, it can cover different time windows and capture the dynamic correlation of timing features such as temperature gradient changes.

[0178] After TCN processing, the output feature vector is reduced in dimension through a fully connected layer and converted into a low-dimensional time series vector.

[0179] Furthermore, the feature vectors output by the multimodal feature branch, the unimodal feature branch, and the temporal feature branch undergo dynamic scaling and gating. The weights are calculated using a sigmoid function, and the feature vectors of each modality are scaled to strengthen the correlation between features from different modalities. The dynamically scaled feature vectors are concatenated to produce a high-dimensional fused vector. This fused vector undergoes dimensionality reduction through a fully connected layer, and nonlinearity is introduced using the ReLU activation function to further extract deeper information from the fused features.

[0180] Optionally, mean square error is suitable for regression tasks, but this embodiment is a classification task, which requires probabilistic output rather than numerical regression, so it is not applicable; binary cross entropy is only suitable for binary classification, and this embodiment is clearly a multi-classification scenario and cannot be used directly. Therefore, this embodiment selects the multi-classification cross entropy loss function as the loss function for model training in this embodiment:

[0181]

[0182] Loss function L cls By calculating the true label y for each sample i,c and predicted probability Logarithmic operations and weighted summation are performed to force the model to predict the probability of the correct category as close to 1 as possible and the prediction probability of the wrong category as close to 0 as possible, thereby improving the discrimination ability in multi-category scenarios.

[0183] The minimum gradient descent strategy is the most commonly used parameter optimization method in deep learning and is suitable for the complex multimodal fusion model in this embodiment. The model includes multiple branches such as multimode fiber feature extraction, single-mode fiber feature processing, and time series feature analysis. The parameter scale is large. Gradient descent can iteratively calculate gradients and update weights, gradually reducing the loss function value and adapting to the nonlinearity and high-dimensional parameter space of the model. Gradient descent can uniformly handle the feature interactions of different modes, ensuring that the feature extractors of each branch (such as lightweight convolution, void convolution, and time convolution network) co-evolve during the optimization process, thereby improving the overall feature fusion effect.

[0184] The beneficial effects of the present invention are as follows:

[0185] By cutting a shallow concave groove at the edge of the road and fixing a flat ribbon optical fiber unit in the groove with polyurea, the present invention constructs a multifunctional distributed optical fiber road sensor that is integrated with the road surface, has high sensitivity, is compact, simple to install, convenient and fast to maintain and repair, and has low cost. By using zero-crossing point and noise thresholds, the normal time domain signal interval determination is optimized, improving the effect of subsequent filtering processing. By using multimodal fusion technology to construct a road condition recognition data model, the recognition accuracy of multiple road abnormalities and the accuracy of abnormal event positioning are improved.

[0186] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0187] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for on-site production of distributed optical fiber road sensors and road risk monitoring and early warning, characterized in that: include: A non-metallic flat ribbon optical fiber unit with 5 or more cores and a thickness of 1.0 mm is manufactured using epoxy glass fiber resin; The non-metallic flat ribbon optical fiber unit includes: 1-core multimode optical fiber and 4-core single-mode optical fibers; Use a diamond blade to cut a shallow concave groove with a width and depth of 4 mm at a distance of 50 mm from the edge of the target road surface; A 1mm thick layer of flexible polyurea is sprayed on the bottom of the concave shallow groove, and the non-metallic flat ribbon optical fiber unit is evenly and flatly fixed on the polyurea surface at the bottom of the concave shallow groove. Rigid polyurea with a temperature resistance of 200°C is injected into the concave shallow groove until it is flush with the road surface. After the polyurea is cured, a distributed optical fiber road sensor is obtained. Using the distributed optical fiber road sensor to collect time and space data on the target road surface without blind spots to obtain road surface status data; The road surface state data is divided into intervals using a zero-crossing point and a noise threshold to obtain a road surface event time domain signal and a normal time domain signal; Performing Fourier transformation and filtering on the road surface event time domain signal according to the normal time domain signal to obtain a road surface abnormal event characteristic signal model to establish a characteristic event database; Performing multimodal feature extraction on the feature event database to obtain multimodal fusion analysis data; The multimodal fusion analysis data is input into a pre-trained road condition recognition model for training to obtain various road abnormality recognition results.

2. The method for on-site production of a distributed optical fiber road sensor and road risk monitoring and early warning according to claim 1 is characterized in that: A 1mm thick layer of flexible polyurea is sprayed on the bottom of the concave shallow groove, and the non-metallic flat ribbon optical fiber unit is evenly and flatly fixed on the polyurea surface at the bottom of the concave shallow groove. Rigid polyurea with a temperature resistance of 200°C is injected into the concave shallow groove until it is flush with the road surface. After the polyurea is cured, a distributed optical fiber road sensor is obtained, including: A bracket is made, two polyurea spray guns are fixed to the front and rear ends of the bracket, and a positioner is installed in front of the nozzle of the polyurea spray gun at the rear end to obtain an optical fiber laying device; the width of the positioner is 0.5 mm less than the width of the concave shallow groove; Mounting the optical fiber ribbon reel of the flat optical fiber ribbon unit on the bracket; During construction, impurities in the concave shallow groove are cleaned, and a layer of polyurea with a thickness of 1 mm is sprayed on the bottom of the concave shallow groove using the polyurea spray gun at the front end. The flat ribbon optical fiber unit is placed at the bottom of the concave shallow groove using the positioner, and polyurea is injected into the concave shallow groove using the polyurea spray gun at the rear end until it is flush with the road surface. After the polyurea is cured, a distributed optical fiber road sensor is obtained.

3. The method for on-site fabrication of a distributed optical fiber road sensor and road risk monitoring and early warning according to claim 1, characterized in that: The distributed optical fiber road sensor is used to collect time and space data of the target road surface without blind spots to obtain road surface status data, including: An optical time domain reflectometer and a multi-channel synchronous acquisition card are respectively connected to the head and tail of the distributed optical fiber road sensor; Using the optical time domain reflectometer, a laser with a wavelength of 1550 nm, a pulse width of 50 ns to 100 ns, and a sampling frequency of 3 kHz is emitted into the single-mode optical fiber, and a laser with a wavelength less than 1000 nm, a pulse width greater than 100 ns, and a sampling frequency of 1.5 kHz is emitted into the multimode optical fiber; Using the multi-channel synchronous acquisition card to synchronously demodulate the optical signals output by the single-mode optical fiber and the multi-mode optical fiber to obtain the road surface state data; When the zero-crossing rate of the road surface state data within a unit time window is greater than a preset threshold value of normal road surface conditions, the sampling frequencies of the single-mode optical fiber and the multi-mode optical fiber are adjusted to 5 kHz and 2.5 kHz, respectively.

4. The method for on-site fabrication of a distributed optical fiber road sensor and road risk monitoring and early warning according to claim 1, characterized in that: The road surface state data is divided into intervals using a zero-crossing point and a noise threshold to obtain a road surface event time domain signal and a normal time domain signal, including: Extracting an interval greater than the noise threshold from the road surface state data to obtain an initial interval; Extracting zero-crossing points in the road surface state data to obtain a zero-crossing point set; Determining the interval range of the two zero-crossing points closest to the initial interval as a road event time domain interval based on the zero-crossing point set, and determining the interval range of the road state data excluding the road event time domain interval as a background time domain interval; The road surface state data is divided into intervals according to the road surface event time domain interval and the background time domain interval to obtain the road surface event time domain signal and the normal time domain signal.

5. The method for on-site fabrication of a distributed optical fiber road sensor and road risk monitoring and early warning according to claim 1, characterized in that: Performing Fourier transformation and filtering on the road event time domain signal according to the normal time domain signal to obtain a road abnormal event characteristic signal model to establish a characteristic event database, including: Performing frequency domain conversion on the road event time domain signal and the normal time domain signal using fast Fourier transform to obtain a road event frequency domain signal and a background frequency domain signal; A difference calculation is performed on the road event frequency domain signal and the background frequency domain signal to obtain the characteristic event database.

6. The method for on-site fabrication of a distributed optical fiber road sensor and road risk monitoring and early warning according to claim 5, characterized in that: Perform multimodal feature extraction on the feature event database to obtain multimodal fusion analysis data, including: The high-frequency energy distribution in each time window of the characteristic event database is extracted to obtain the STFT time-frequency vector; the expression of the STFT time-frequency vector is: X stft =[E 100-200 ,E 200-300 ,E 300-500 ,Peak_freq,Energy_std]; Among them, X stft is the STFT time-frequency vector; E a-b Represents the normalized energy in the frequency band [a, b]; Peak_freq represents the main frequency of the impulse signal; Energy_std represents the standard deviation of the energy in the frequency band from 100Hz to 500Hz; Perform wavelet decomposition on the characteristic event database and calculate the energy entropy at each scale to obtain a wavelet energy entropy vector; the expression of the wavelet energy entropy vector is: wavelet =[H1,H2,H3]; where X wavelet is the wavelet energy entropy vector; H1, H2, and H3 are the energy entropies of the 1st, 2nd, and 3rd layers of the wavelet decomposition of the characteristic event database respectively; Extract the time domain instantaneous characteristics of the road event frequency domain signal to obtain an instantaneous feature vector; the expression of the instantaneous feature vector is: transient =[t rise ,t duration ]; where X transient is the instantaneous eigenvector; t rise The time it takes for the signal to rise from 10% to 90% of its peak value; t duration is the duration of the impact signal; The STFT time-frequency vector, the wavelet energy entropy vector, and the instantaneous feature vector are fused to obtain a mechanical impact vector.

7. The method for on-site fabrication of a distributed optical fiber road sensor and road risk monitoring and early warning according to claim 6, characterized in that: Perform multimodal feature extraction on the feature event database to obtain multimodal fusion analysis data, including: The energy proportion and modal frequency of the low-frequency signal in the characteristic event database are calculated to obtain a low-frequency energy eigenvector; the expression of the low-frequency energy eigenvector is: X lowfreq =[E low ,f mode1 ,f mode2 ]; Among them, X lowfreq is the low-frequency energy eigenvector; E low is the energy proportion of the frequency band less than 100 Hz; f mode1 is the first-order vibration mode frequency; f mode2 is the second-order vibration mode frequency; Calculate the variance and skewness of the characteristic event database to obtain the strain-frequency coupling characteristic vector; the expression of the strain-frequency coupling characteristic vector is: X strain_freq =[Var_f,Skew_f]; where X strain_freq is the strain-frequency coupling eigenvector; Var_f is the variance of the signal frequency; Skew_f is the skewness of the signal frequency; Extract the gradient change of the temperature channel of the characteristic event database of the two time periods to obtain the temperature anomaly feature vector; the expression of the temperature anomaly feature vector is: temp =[ΔT1,ΔT2]; where X temp is the temperature anomaly characteristic vector; ΔT1 and ΔT2 are the temperature change rates in the first time period and the second time period respectively; the second time period is greater than the first time period; The low-frequency energy eigenvector, the strain-frequency coupling eigenvector, and the temperature anomaly eigenvector are fused to obtain a structural deformation vector.

8. The method for on-site fabrication of a distributed optical fiber road sensor and road risk monitoring and early warning according to claim 7, characterized in that: Performing multimodal feature extraction on the feature event database to obtain multimodal fusion analysis data also includes: The energy entropy of the high frequency band and the power spectrum density of the low frequency band of the characteristic event database are extracted to obtain a first environmental feature vector; the expression of the first environmental feature vector is: X1=[H 1_high ,PSD 1_low ]; Wherein, X1 is the first environmental feature vector; H 1_high Energy entropy from 50Hz to 200Hz; PSD 1_low is the power spectrum density value of the frequency band less than 30Hz; The low-frequency energy proportion and high-frequency pulse frequency of the characteristic event database and the signal duration and low-frequency energy rising slope of the road event frequency domain signal are extracted to obtain a second environmental feature vector; the expression of the second environmental feature vector is: X2=[E 2_low ,f impulse ,t duration ,Slope_E]; Wherein, X2 is the second environment feature vector; E 2_low is the energy proportion of the frequency band less than 20Hz; f impulse is the pulse frequency from 50Hz to 200Hz; Slope_E is the rate of change of energy in the frequency band less than 20Hz; splicing the mechanical impact vector, the structural deformation vector, the first environmental feature vector, and the second environmental feature vector to obtain an auxiliary feature matrix; Converting the characteristic event database corresponding to the multimode optical fiber and the single-mode optical fiber into a multimode optical fiber time-frequency diagram and a single-mode optical fiber time-frequency diagram; The multimode optical fiber time-frequency diagram, the single-mode optical fiber time-frequency diagram, and the auxiliary feature matrix are fused to obtain the multimodal fusion analysis data.

9. The method for on-site fabrication of a distributed optical fiber road sensor and road risk monitoring and early warning according to claim 8, characterized in that: The training process of the road condition recognition model includes: Pre-collecting the multimodal fusion analysis data and adding real road condition labels to the multimodal fusion analysis data; the real road condition labels include: collision, vehicle rollover, tire blowout, road collapse, sudden heavy rainfall, road icing, falling rocks, mudslide, fallen objects, and explosions; Build the basic model architecture; Using the basic model architecture, the multimode optical fiber time-frequency graph is sequentially subjected to lightweight convolution, hole convolution processing, and group convolution to obtain a multimode optical fiber characteristic graph; The spatial attention gating mechanism of the basic model architecture is used to perform maximum pooling, average pooling, linear transformation, and Sigmoid function activation on the multimode fiber feature map to obtain an attention weight; Multiplying the attention weight and the multimode fiber feature map and adjusting the size to a preset fixed size through an adaptive pooling operation to obtain a multimode feature; Using the basic model architecture, lightweight convolution and time dimension pooling are performed on the single-mode optical fiber time-frequency map to obtain a single-mode optical fiber feature map; Performing modal interaction on the single-mode optical fiber feature map and adjusting the size to the fixed size through an adaptive pooling operation to obtain a single-mode feature; Using the temporal convolutional network of the basic model architecture to perform causal convolution processing on the auxiliary feature matrix and perform dimensionality reduction through a fully connected layer to obtain a time series vector; The basic model architecture is used to perform dynamic scaling gating processing, feature splicing and dimensionality reduction processing on the multi-mode features, the single-mode features and the time series vector respectively to obtain fused features.

10. The method for on-site fabrication of a distributed optical fiber road sensor and road risk monitoring and early warning according to claim 9, characterized in that: The training process of the road condition recognition model further includes: Utilizing the Softmax activation function to calculate the road condition probability distribution of the fused features, and obtaining a network prediction result; The loss function is calculated using the network prediction result and the real road condition label to obtain the model loss value; the expression of the loss function is: Among them, L cls is the model loss value; M is the number of batch training samples; N is the total number of road condition categories; y i,c is the true label of the i-th sample, which is 1 if it belongs to category c, otherwise it is 0; is the predicted probability that the i-th sample belongs to category c in the network prediction result; The basic model architecture is parameter optimized and iterated using the model loss value according to the minimum gradient descent strategy to obtain the trained road condition recognition model.

Images