Iot intelligent road network safety monitoring system based on distributed optical fiber sensing

Abstract

The application discloses an Internet of Things intelligent road network safety monitoring system based on distributed optical fiber sensing, which comprises a sensing module, a transmission module, a processing module and a disease development prediction module connected in sequence. The sensing module comprises a sensing optical fiber network, a light source submodule and a photoelectric detection submodule. The transmission module is used for uploading the electrical signal to the processing module. The processing module is used for processing the electrical signal, and performing disease classification identification and health state evaluation. The disease development prediction module is used for predicting the disease development and providing a maintenance scheme. The application fuses a lightweight deep learning model, a hybrid neural network algorithm and a Monte Carlo simulation technology to construct the Internet of Things intelligent road network safety monitoring system based on distributed optical fiber sensing, realizes efficient processing of sensing data, accurate identification and prediction of diseases and cost benefit optimization of the maintenance scheme, and solves the pain points of the prior art.

Description

Technical Field

[0001] This invention relates to the field of fiber optic sensing road network monitoring technology, and in particular to an Internet of Things (IoT) intelligent road network safety monitoring system based on distributed fiber optic sensing. Background Technology

[0002] Existing road network monitoring systems largely rely on point sensors (such as strain gauges and accelerometers), which suffer from limited monitoring range, high deployment costs, and weak anti-interference capabilities, making them unsuitable for application scenarios involving winding mountain roads and complex environments (high temperature, high humidity, and strong vibration). While distributed fiber optic sensing technology can achieve large-scale continuous monitoring, it faces bottlenecks such as high-dimensional sensor data, excessive redundant information, and difficulty in feature extraction. Furthermore, traditional monitoring systems lack the ability to accurately predict the development of road damage and quantitatively evaluate maintenance plans, leading to delayed maintenance decisions and significant resource waste. Summary of the Invention

[0003] The purpose of this invention is to address the aforementioned problems by providing an IoT-based intelligent road network safety monitoring system that integrates lightweight deep learning models, hybrid neural network algorithms, and Monte Carlo simulation technology. This system enables efficient processing of sensor data, accurate identification and prediction of road defects, and cost-effective optimization of maintenance plans, thus resolving the pain points of existing technologies.

[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: According to one aspect of the present invention, an Internet of Things (IoT) smart road network safety monitoring system based on distributed optical fiber sensing is provided, comprising a sensing module, a transmission module, a processing module and a disease development prediction module connected in sequence. The sensing module includes a sensing fiber optic network deployed in the road to be monitored, a light source submodule for injecting detection light signals into the sensing fiber optics, and a photoelectric detection submodule for receiving backscattered light signals. The transmission module is used to upload the electrical signal output by the photoelectric detection submodule to the processing module. The processing module is used to process the electrical signal and perform disease classification and identification and health status assessment. The disease development prediction module is used to predict disease development and provide maintenance solutions.

[0005] Preferably, the sensing fiber optic network includes single-mode fiber, graphene-coated fiber, and stretchable fiber wrapped with PDMS elastomer.

[0006] Preferably, the transmission module includes a dual-link redundant transmission structure and a signal relay device, wherein the dual-link redundant transmission structure includes a 5G wireless network and a wired network.

[0007] Preferably, the processing module includes a feature extraction submodule, a disease identification submodule, and a state assessment and error correction submodule; The feature extraction submodule is based on the MobileNetV3 lightweight convolutional neural network model and is used to extract features from real-time data. The disease identification submodule adopts the LSTM-Transformer hybrid neural network model to achieve accurate classification and identification of road network diseases; The state assessment and error correction submodule is based on a BP neural network and is used for quantitative assessment of the overall health status of the road network and error correction of the LSTM-Transformer hybrid network defect identification results.

[0008] Preferably, the MobileNetV3 lightweight convolutional neural network model includes depthwise separable convolutional layers, an attention mechanism, pooling operation layers, and fully connected layers; The depthwise separable convolutional layer includes depthwise convolution and pointwise convolution. The depthwise convolution is used to allocate a convolutional kernel to each input channel to complete the spatial feature extraction within the channel. The pointwise convolution is used to complete the feature fusion and dimension adjustment between different channels. The attention mechanism includes an SE channel attention mechanism and a temporal attention module; The pooling operation layer includes an input size sensing unit, a pooling parameter dynamic calculation unit, and an adaptive averaging calculation unit; The fully connected layer includes an input feature receiving unit, a first linear mapping layer, a lightweight activation layer, a second linear mapping layer, and an output feature output unit.

[0009] Preferably, the LSTM-Transformer hybrid neural network model includes a feature fusion layer, a temporal coding layer, a global coding layer, a feature aggregation layer, and a classification output layer.

[0010] Preferably, the BP neural network includes a multidimensional feature input layer, a hidden layer, and a quantization result output layer.

[0011] Preferably, the disease development prediction module includes a Monte Carlo simulation unit. Monte Carlo simulation constructs a probability model, performs a large number of independent random samplings on the random variables in the model, substitutes the sampling results into the model for iterative calculation, and finally obtains the probability distribution and confidence interval of the results through statistical analysis.

[0012] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This invention achieves coordinated monitoring of multiple parameters, including vibration, temperature, and strain, by simultaneously acquiring Rayleigh scattering, Raman scattering, and Brillouin scattering signals through a single sensing fiber, overcoming the limitations of single-scattering technologies. A processing module extracts features from the data and uses these features for disease identification and classification, while simultaneously assessing road network health and correcting model errors, improving processing efficiency. It is adaptable to edge node deployment, has strong anti-interference capabilities, and can meet the continuous monitoring needs of complex road conditions such as mountain roads and expressways. Finally, a disease development prediction module enables quantitative prediction of disease development and cost-effectiveness optimization of maintenance plans, solving the problems of lagging traditional maintenance decisions and resource waste. Attached Figure Description

[0013] Figure 1 This is a functional structure diagram of the present invention; Figure 2 This is a functional structure diagram of the MobileNetV3 lightweight convolutional neural network model of the present invention; Figure 3 This is a functional structure diagram of the LSTM-Transformer hybrid neural network model of the present invention; Figure 4 This is a schematic diagram of the functional structure of the BP neural network of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the invention, and these aspects of the invention can be achieved even without these specific details.

[0015] Please see Figures 1-4 This invention provides an IoT-based intelligent road network safety monitoring system based on distributed optical fiber sensing, the technical solution of which is as follows: The IoT-based intelligent road network safety monitoring system based on distributed optical fiber sensing includes a sensing module, a transmission module, a processing module, a disease development prediction module, and an execution module controlled by the disease development prediction module, connected in sequence. The sensing module includes a sensing optical fiber network deployed in the road to be monitored, a light source submodule for injecting detection light signals into the sensing optical fiber, and a photoelectric detection submodule for receiving backscattered light signals. The sensing optical fiber network includes single-mode optical fiber, stretchable optical fiber wrapped with graphene-coated fiber and PDMS elastomer. The single-mode optical fiber is G.655 bend-resistant single-mode fiber. Due to road construction and long-term geological activities, optical fibers may be subjected to compression and bending. Bending-resistant single-mode fiber can significantly reduce the additional signal loss caused by micro-bending, ensuring the stability of long-distance signal transmission. Graphene-coated optical fiber can be used in high-temperature key monitoring areas, such as bridge deck pavement, tunnel fire early warning points, and road sections near thermal pipelines. The graphene coating provides better heat resistance and thermal conductivity, protecting the optical fiber structure and enabling the optical fiber to respond quickly to changes in ambient temperature, ensuring the reliability and accuracy of temperature measurement. PDMS elastomer-coated stretchable optical fibers are used in areas expected to experience large deformations or active joints, such as near bridge expansion joints, settlement sections of soft soil subgrades, and potential slip zones on slopes. Elastomer-coated stretchable optical fibers significantly increase the effective strain monitoring range of the fiber by up to 300%. This allows the system to monitor not only micro-strain but also large, non-destructive structural displacements, expanding its application boundaries. Deployment methods include direct burial at a depth of at least 50 cm at the bottom of the subgrade and binding fixation inside bridge box girders. Single-mode fiber is used as the main fiber in most road sections, graphene-coated fiber is used in high-temperature sections, and PDMS elastomer-coated stretchable optical fibers are used in bridge expansion joints and settlement sections of soft soil subgrades. The optical fiber network is used to monitor strain, settlement, and temperature gradients within the subgrade. Buried within the subgrade structure, it can detect uneven settlement originating from the foundation earlier, enabling early warning. By placing optical fibers below the main load-bearing layer and the freeze-thaw effect layer, it is possible to stably sense the deformation of the roadbed while avoiding damage caused by the direct impact of upper road construction and heavy traffic. The fibers are also bound and fixed inside the bridge box girder to monitor the strain distribution, vibration characteristics, and internal temperature field of key load-bearing components.

[0016] The light source submodule consists of a high-power pulsed laser and a wavelength-tunable continuous-wave laser, both used to generate probe light signals injected into the sensing fiber. The high-power pulsed laser emits the probe light signal to generate Rayleigh and Brillouin scattering. The wavelength-tunable continuous-wave laser is used for OFDR. The photodetector submodule receives extremely weak backscattered light, such as Rayleigh, Raman, and Brillouin scattered light, returning from the sensing fiber and converts it into an electrical signal for subsequent demodulation. The photodetector submodule includes a balanced detector, which greatly suppresses common-mode noise, improves the signal-to-noise ratio, and increases photoelectric conversion efficiency. The balanced detector can detect extremely weak light signals, enabling long-distance, high-precision measurements.

[0017] The transmission module uploads the electrical signals output by the photoelectric detection submodule to the processing module. Specifically, the transmission module adopts a dual-link redundant transmission structure consisting of a 5G wireless network and a wired network, and deploys signal relay equipment when the monitoring distance exceeds 50 kilometers. Through heterogeneous redundancy and automatic switching mechanisms between wired and wireless networks, the risk of single network failure is effectively addressed. The 5G wireless network solves the problem of coverage blind spots in purely wired networks, supports more complex road network environments and mobile application scenarios, and expands the system's applicable boundaries. The intelligent signal relay equipment solves the problem of long-distance signal attenuation in distributed sensing, achieving seamless monitoring over ultra-long distances.

[0018] The processing module includes an FPGA processor, which includes a feature extraction submodule, a disease identification submodule, and a condition assessment and error correction submodule.

[0019] Due to the high dimensionality, high redundancy, and strong temporal correlation of distributed fiber optic sensor data, and the need for road network monitoring to address environmental noise and sensor data fragments of varying lengths, the feature extraction submodule employs the MobileNetV3 model to extract real-time features from the data collected by the distributed fiber optic sensor system. This outputs candidate feature vectors for defects, providing a foundation for subsequent defect identification and resolving the issues of excessive latency and bandwidth consumption associated with traditional centralized data processing.

[0020] Specifically, the MobileNetV3 model includes depthwise separable convolutional layers, attention mechanisms, pooling operation layers, and fully connected layers.

[0021] Depthwise separable convolutional layers include depthwise convolution and pointwise convolution. Depthwise convolution allocates a separate kernel to each input channel, extracting spatial features within each channel without fusing features between channels, significantly reducing the computational cost of convolution. Pointwise convolution uses a 1×1 kernel to fuse features between different channels and adjust dimensions.

[0022] The attention mechanism includes a channel-specific attention mechanism (SE) and a temporal attention module. The SE channel-specific attention mechanism achieves channel-level feature weight allocation through two steps: squeezing and activation. The squeezing step performs global average pooling on the feature map after depthwise convolution, compressing the spatial features of each channel into a single global feature value. The activation step learns the importance weights of different channels through two fully connected layers, weighting and enhancing the feature map to highlight effective features and suppress ineffective features, thereby improving the model's ability to capture key disease features.

[0023] To address the temporal correlation of vibration and strain signals in road network sensor data, the temporal attention module learns temporal weights from consecutive frames of sensor data. This strengthens the temporal correlation of defect features while suppressing interference from environmental noise such as vehicle traffic and natural vibrations, effectively solving the problem of environmental noise masking defect features under complex road conditions. Furthermore, for consecutive frames of fiber optic sensor data, the temporal attention module assigns high weights to defect-related temporal features and low weights to environmental noise temporal features through three steps: temporal feature encoding, similarity calculation, and weight normalization. Ultimately, this achieves the goal of strengthening effective temporal features and suppressing ineffective temporal noise.

[0024] Specifically, the temporal attention module consists of a temporal feature extraction layer, an attention weight calculation layer, and a weighted fusion layer.

[0025] The input data for the temporal feature extraction layer is a sequence of single-channel / multi-channel feature maps after depthwise separable convolution and SE channel attention processing. The temporal feature extraction layer uses 1×1 lightweight convolution and temporal pooling to extract temporal features. The 1×1 convolution compresses the dimensionality of the feature maps, while temporal pooling captures the basic temporal correlation features between consecutive frames, outputting a temporal feature matrix. This transforms the spatial feature maps into a temporal-feature two-dimensional matrix, providing a temporal feature foundation for subsequent weight calculations.

[0026] The attention weight calculation layer generates temporal attention weights through self-attention similarity calculation and Softmax normalization. Specifically, through self-attention similarity calculation, the cosine similarity between each temporal frame feature in the temporal feature matrix and all other frame features is calculated to measure the temporal correlation strength between frames, generating a similarity matrix. Then, through weight normalization, each row of the similarity matrix is ​​normalized using the Softmax function, converting the correlation strength into attention weights between 0 and 1, generating an attention weight matrix that ensures the weights of each row sum to 1, guaranteeing the rationality of weight allocation. High weights are assigned to disease feature frames with strong temporal correlation, and low weights are assigned to noise feature frames with weak temporal correlation, achieving feature selection in the temporal dimension.

[0027] The weighted fusion layer performs matrix multiplication between the attention weight matrix and the original temporal feature matrix to obtain a weighted temporal feature matrix. Then, by concatenating the time dimensions, the weighted temporal features are transformed back to a feature map format matching the original model output. Attention weights are applied to the original temporal features to enhance high-weighted disease temporal features and suppress low-weighted noisy temporal features. The final output is a feature map fused with temporal attention, providing a high-quality temporal feature foundation for subsequent disease identification using the LSTM-Transformer hybrid network.

[0028] The pooling operation layer includes an input size-aware unit, a pooling parameter dynamic calculation unit, and an adaptive averaging calculation unit.

[0029] The input size sensing unit identifies the spatial dimension parameters of the input feature map in real time. It receives feature map input from the temporal attention module and extracts its height, width, and number of channels, providing a foundation for subsequent dynamic calculations. The input size sensing unit can directly sense the dimension parameters of any size feature map formed after convolution or attention processing of road network sensing data, without prior standardization, thus solving the problem of inconsistent data sizes for different monitoring sections of mountainous highways.

[0030] The pooling parameter dynamic calculation unit automatically calculates the pooling kernel size and sliding step size based on the preset output dimension and the results of the input size sensing unit. By presetting the output dimension to 1×1×C, the pooling kernel size and sliding step size are calculated, achieving spatial dimension compression in a single pooling operation.

[0031] The adaptive averaging unit uses the calculated pooling kernel and stride to perform global averaging sampling on the input feature map in spatial dimension, outputting a feature vector with a fixed 1×1×C dimension. Specifically, a dynamically calculated pooling kernel covers the entire spatial region of a single feature map, and the arithmetic mean of all pixels within the pooling kernel is taken. Each channel is calculated independently, and finally, each channel outputs an average value, forming a feature value set in channel dimension. This preserves the global information of the features while completing the standardization and compression of the spatial dimension.

[0032] Adaptive average pooling can automatically adjust the pooling kernel size and stride according to the size of the input feature map, outputting a fixed-dimensional feature vector without requiring additional data pruning or zero-padding. This improves the model's generalization ability and adapts to road network monitoring scenarios where distributed fiber optic sensor data segments have varying lengths and dimensions, solving the problem that traditional fixed-kernel pooling layers cannot flexibly adapt to different input feature sizes.

[0033] The fully connected layer includes an input feature receiving unit, a first linear mapping layer, a lightweight activation layer, a second linear mapping layer, and an output feature output unit.

[0034] The input feature receiving unit receives fixed-dimensional feature vectors from the adaptive average pooling layer, flattens the three-dimensional feature vectors into one-dimensional vectors, and prepares them for subsequent linear mapping.

[0035] The first linear mapping layer performs dimensionality reduction and compression of feature vectors through a trainable weight matrix, filtering out redundant features and retaining core features related to road network defects.

[0036] The lightweight activation layer performs a nonlinear transformation on the output of the first linear mapping layer. The Hard-Swish activation function enhances the discriminative power of defect features, especially weak defect features such as microcracks and slight subsidence. The smooth nonlinear characteristics of the Hard-Swish activation function can better capture weak defect features of the road network.

[0037] The second linear mapping layer maps the dimensionality-reduced and activated feature vectors to fixed-dimensional disease candidate feature vectors, which are then directly connected to the feature fusion layer of the subsequent LSTM-Transformer hybrid network.

[0038] The output feature unit takes the output of the second linear mapping layer as the disease candidate feature vector and directly passes it to the feature fusion layer of the LSTM-Transformer hybrid network. This ensures that the feature vector dimension perfectly matches that of the subsequent network, without any dimension transformation loss.

[0039] To address the issue that road network defects exhibit both temporal correlation and global correlation, the defect identification submodule constructs an LSTM-Transformer hybrid network to achieve accurate classification and identification of road network defects.

[0040] Specifically, the LSTM-Transformer hybrid network includes a feature fusion layer, a temporal coding layer, a global coding layer, a feature aggregation layer, and a classification output layer.

[0041] The feature fusion layer fuses the disease candidate feature vectors output by the MobileNetV3 model with the original temporal feature vectors from the fiber optic sensing data, forming a unified input feature sequence and addressing the issue of insufficient single feature dimensions. The feature vectors output by the MobileNetV3 model and the original temporal feature vectors from the sampled fiber optic sensing data are concatenated using dimensionality concatenation and batch normalization (BN). The resulting fused feature vector is then eliminated through a BN layer to eliminate dimensionality differences, outputting the feature sequence.

[0042] The temporal coding layer is a Bi-LSTM local temporal coding layer, which encodes the local short-term temporal dependencies of the sensor data to capture short-term variation patterns of disease features. Its network type is a bidirectional LSTM (Bi-LSTM), which includes a forward LSTM and a backward LSTM, capable of simultaneously capturing either forward or backward temporal dependencies. The forward LSTM encodes the feature sequence and outputs the forward hidden state. The backward LSTM encodes the feature sequence and outputs the backward hidden state. The forward and backward hidden states are concatenated dimensionally to obtain the bidirectional hidden state.

[0043] The global encoding layer is a lightweight Transformer encoding layer that captures long-term temporal global dependencies in sensor data and identifies long-term trends in disease development. The bidirectional hidden states output by the Bi-LSTM are used as input to the Transformer. Global correlation weights between frames are calculated through multi-head self-attention, and after an FFN linear transformation, the globally encoded features are output.

[0044] The feature aggregation layer compresses the temporal feature sequence into a fixed-dimensional global feature vector, preparing for final classification. Temporal-dimensional global average pooling averages the features across all temporal frames of the feature sequence to obtain the global feature vector.

[0045] The classification output layer maps the global feature vector to the probability distribution of road network defect types and outputs the final identification result.

[0046] By taking the candidate feature vectors of road network defects output by MobileNetV3, the system can achieve accurate classification and identification of road network defects. It combines the advantages of LSTM in encoding time-series features and the ability of Transformer to capture long-distance dependencies. At the same time, it has been customized in a lightweight manner to address the time-series characteristics of distributed fiber optic sensing data and the computing power limitations of edge nodes.

[0047] The Condition Assessment and Error Correction submodule is used for quantitative assessment of the overall health status of the road network and error correction of the LSTM-Transformer hybrid network defect identification results. This submodule employs a BP neural network to connect the defect identification results from the LSTM-Transformer to perform quantitative assessment of the overall health status of the road network and error correction of the defect identification results, providing a quantitative basis for subsequent defect prediction and maintenance decisions.

[0048] Specifically, a BP neural network includes a multidimensional feature input layer, a hidden layer, and a quantization result output layer.

[0049] The multi-dimensional feature input layer receives standardized multi-dimensional feature parameters and passes them directly to the hidden layer.

[0050] The hidden layer consists of two cascaded layers. The first hidden layer performs a first nonlinear mapping on the dimensional features of the input layer, outputting an intermediate feature vector. The second hidden layer performs a second nonlinear dimensionality reduction mapping on the intermediate feature vector, outputting a core feature vector, providing a high-dimensional foundation for the quantization calculation of the output layer.

[0051] The quantization result output layer uses a single output node, distinguishing between the health status assessment function and the error correction function through the semantic definition of the output values. The output layer outputs the quantization results using the following formula: in, The final quantification result (health score or correction coefficient) of the output layer; This is the output layer weight matrix; This is the output layer bias term, a single value used to fine-tune the output baseline; This is the core feature vector output by the second hidden layer.

[0052] During the network training phase in the output layer, optimization is achieved through backpropagation. and , to output The goal is to approximate the human-annotated true values ​​(true health scores or true error correction coefficients) as closely as possible. During the inference phase, the pre-trained values ​​are fixed. and The result is directly calculated and output using the formula.

[0053] When assessing the health status of the road network, the input layer receives all 12-dimensional features, which are then mapped nonlinearly through two hidden layers to output a 32-dimensional core feature vector. The trained weight matrix W performs a weighted summation and biasing on the 32-dimensional feature vectors. The weight values ​​reflect the degree of influence of each feature on the health status. The output y is limited to the range of 0-100 and is directly used as the road network health score. For example, 85≤y≤100 is excellent; 70≤y<85 is good; 50≤y<70 is average; and y<50 is poor.

[0054] When correcting errors in disease identification results, the input layer selects only 3-dimensional core features. After mapping through two hidden layers, a 32-dimensional core feature vector V is output. The trained weight matrix W weights the 32-dimensional feature vector, correcting coefficients for identification results with high error risk. The output y is limited to the range of 0.8-1.2 and used as an error correction coefficient to adjust the recognition probability of the LSTM-Transformer. The corrected recognition probability = the original recognition probability of the LSTM-Transformer × y.

[0055] The disease development prediction module uses a Monte Carlo simulation engine for disease development prediction and cost-benefit analysis of maintenance plans.

[0056] Specifically, Monte Carlo simulation constructs a probabilistic model, performs extensive independent random sampling of the random variables in the model, substitutes the sampling results into the model for iterative calculation, and finally obtains the probability distribution and confidence interval of the results through statistical analysis. This quantifies the uncertainty of the model and solves the problem that traditional deterministic models cannot consider the random fluctuations of multiple factors.

[0057] Monte Carlo simulation for road network disease development prediction aims to forecast medium- to long-term trends. Targeting two core road network diseases—crack propagation and subgrade settlement—it outputs the time points and confidence intervals for diseases to reach a critical safety state, providing a scientific basis for determining maintenance timing. Specifically, it includes the following steps: S1. Constructing a probability model and dynamic equations for disease development: Determine the target disease and critical state for prediction, and construct dynamic equations. The dynamic equations include crack propagation and subgrade settlement. Crack propagation is described using the Paris equation, which describes the relationship between the crack propagation rate and the stress intensity factor. The formula is as follows: in, This represents the crack propagation rate. , These are constants related to road materials; This is the stress intensity factor.

[0058] The equation for the growth of subgrade settlement with time, traffic volume, and precipitation is as follows: in, for Accumulated settlement over time. This is the final settlement. The consolidation coefficient is . For traffic volume Precipitation intensity The correction function.

[0059] The core random variables affecting disease development were extracted from the dynamic equations. Based on more than 5 years of historical monitoring data of mountain roads, the probability distribution type and parameters of each variable were determined by statistical fitting.

[0060] S2. Large-scale independent sampling of random variables: Based on the probability distribution of each random variable fitted in S1, 10,000 sets of independent random sampling are performed using a pseudo-random number generation algorithm, resulting in 10,000 sets of random variable combinations. Each set contains specific values ​​for average daily traffic flow, proportion of heavy vehicles, precipitation intensity, temperature range, and initial size of the disease.

[0061] S3. Substitute into the dynamic equation for iterative calculation: Substitute 10,000 sets of random variables into the disease development dynamic equation one by one, and simulate the growth process of disease size over time under each set of variables through iterative calculation with time step until the disease reaches the critical safety state. Record the time value of the disease reaching the critical state under each set of variables.

[0062] For example, if a sample group has an average daily traffic volume of 8,200 vehicles, a heavy vehicle ratio of 37%, an annual precipitation of 1,250 mm, and an initial crack width of 0.8 mm, after substituting into the Paris equation, the time to reach a crack width of 5 mm is calculated to be 3.2 years. If another sample group has an average daily traffic volume of 7,500 vehicles, a heavy vehicle ratio of 32%, an annual precipitation of 1,100 mm, and an initial crack width of 0.8 mm, the critical time is calculated to be 3.8 years.

[0063] S4. Statistical analysis of results and output of prediction conclusions: Statistical analysis is performed on 10,000 sets of critical time values ​​to calculate their probability distribution, mean, and quantiles. Disease development trend curves and critical time confidence intervals are plotted, and the prediction results under multiple confidence levels are finally output.

[0064] The output result is as follows: Disease development trend: At a 95% confidence level, the process of crack width increasing from 0.8 mm to 5 mm is characterized by slow expansion in the first 1.5 years and rapid expansion in the following 1.7 years; Critical time prediction: At a 90% confidence level, the critical time is 2.9-3.5 years; at a 95% confidence level, the critical time is 3.0-3.4 years; at a 99% confidence level, the critical time is 2.8-3.6 years. Recommended timing for core maintenance: Initiate preventative maintenance at 2.5 years. This allows for completion of the work before the disease reaches a critical state, thus avoiding structural damage.

[0065] Based on the critical time points and development trends of road network diseases, Monte Carlo simulation further conducts multi-dimensional cost-benefit quantitative analysis on three types of core road network maintenance schemes: preventive maintenance, restorative maintenance, and emergency maintenance. By constructing a cost-benefit evaluation system and combining random sampling to calculate indicators such as net present value, return on investment, and risk probability of each scheme, the optimal comprehensive maintenance scheme is finally selected to achieve precise allocation of maintenance resources.

[0066] Based on the disease development prediction results, a multi-dimensional maintenance scheme evaluation system was constructed. The cost and benefits of different maintenance schemes were quantified through Monte Carlo simulation to select the optimal scheme. The specific analysis data is shown in the table below (taking a test scenario of a highway in the mountainous area of ​​Guangxi Zhuang Autonomous Region as an example): Note: The data is based on simulation calculations of 1000 samples from Monte Carlo. Baseline parameters: average daily traffic flow of 8000 vehicles, heavy vehicles accounting for 35%, average annual precipitation of 1200mm, and maintenance and construction cycle of 15 days.

[0067] By visually comparing the core indicators of the three schemes using radar charts, the preventative maintenance scheme (crack sealing + roadbed grouting) demonstrates the best overall performance in terms of net present value, return on investment, and risk control, making it the preferred option for this scenario. Furthermore, by combining the disease development trend curves simulated in Monte Carlo simulations, the timing of maintenance implementation can be accurately determined.

[0068] Compared to traditional static maintenance decision-making, this method considers the randomness of disease development and the uncertainty of environmental factors through Monte Carlo simulation, making the cost-benefit analysis of maintenance plans more in line with actual scenarios. This reduces the blindness of maintenance decisions and is expected to reduce total maintenance costs by 20%-30%. Combining edge computing IoT security technology, a collaborative maintenance decision-making security system is constructed: disease prediction data and simulation results uploaded from edge nodes to the backend server are transmitted with end-to-end encryption, and digital signatures and data integrity verification codes are added to the edge nodes to prevent data tampering or forgery, ensuring that maintenance decisions are based on real data. When maintenance plan instructions generated by the backend server are sent to edge nodes and terminal devices, a hierarchical access control mechanism is adopted. Different levels of management personnel have different operating permissions (such as viewing, approving, and executing), and a role-based access control (RBAC) model ensures the security of instruction issuance and execution. A maintenance plan execution data traceability mechanism is established, with edge nodes recording the receipt, execution process, and result data of maintenance instructions, encrypting and storing them, and synchronizing them to the cloud to achieve full-process traceability and prevent unauthorized modification or malicious execution of instructions.

[0069] This invention includes the following steps: Data acquisition: A distributed fiber optic sensor network is deployed along the road network to collect real-time data on the strain, temperature, and vibration of the road surface and subgrade. The data is then transmitted to the intelligent data processing layer via the Internet of Things (IoT) transmission layer. Feature extraction and disease identification: The MobileNetV3 model extracts data features, the LSTM-Transformer hybrid network completes the identification of disease type and level, and the BP neural network performs error correction and road network health status assessment. Disease development prediction: Based on Monte Carlo simulation, combined with assessment results and environmental parameters, predict the disease development trend and critical time nodes; Maintenance plan optimization: The cost-effectiveness of different maintenance plans is analyzed through Monte Carlo simulation, and the optimal maintenance plan is output. Terminal display and decision-making push: The terminal display layer displays road network monitoring data, disease information, prediction results and maintenance plans in real time, and pushes decision-making suggestions to managers.

[0070] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An IoT-based intelligent road network safety monitoring system based on distributed optical fiber sensing, characterized in that, It includes a sensing module, a transmission module, a processing module, and a disease development prediction module connected in sequence; The sensing module includes a sensing fiber optic network deployed in the road to be monitored, a light source submodule for injecting detection light signals into the sensing fiber optics, and a photoelectric detection submodule for receiving backscattered light signals. The transmission module is used to upload the electrical signal output by the photoelectric detection submodule to the processing module. The processing module is used to process the electrical signal and perform disease classification and identification and health status assessment. The disease development prediction module is used to predict disease development and provide maintenance solutions.

2. The IoT intelligent road network safety monitoring system based on distributed optical fiber sensing according to claim 1, characterized in that: The sensing fiber optic network includes single-mode fiber, graphene-coated fiber, and stretchable fiber wrapped with PDMS elastomer.

3. The IoT intelligent road network safety monitoring system based on distributed optical fiber sensing according to claim 1, characterized in that: The transmission module includes a dual-link redundant transmission structure and a signal relay device. The dual-link redundant transmission structure includes a 5G wireless network and a wired network.

4. The IoT intelligent road network safety monitoring system based on distributed optical fiber sensing according to claim 1, characterized in that: The processing module includes a feature extraction submodule, a disease identification submodule, and a state assessment and error correction submodule; The feature extraction submodule is based on the MobileNetV3 lightweight convolutional neural network model and is used to extract features from real-time data. The disease identification submodule adopts the LSTM-Transformer hybrid neural network model to achieve accurate classification and identification of road network diseases; The state assessment and error correction submodule is based on a BP neural network and is used for quantitative assessment of the overall health status of the road network and error correction of the LSTM-Transformer hybrid network defect identification results.

5. The IoT intelligent road network safety monitoring system based on distributed optical fiber sensing according to claim 4, characterized in that: The MobileNetV3 lightweight convolutional neural network model includes depthwise separable convolutional layers, an attention mechanism, pooling operation layers, and fully connected layers; The depthwise separable convolutional layer includes depthwise convolution and pointwise convolution. The depthwise convolution is used to allocate a convolutional kernel to each input channel to complete the spatial feature extraction within the channel. The pointwise convolution is used to complete the feature fusion and dimension adjustment between different channels. The attention mechanism includes an SE channel attention mechanism and a temporal attention module; The pooling operation layer includes an input size sensing unit, a pooling parameter dynamic calculation unit, and an adaptive averaging calculation unit; The fully connected layer includes an input feature receiving unit, a first linear mapping layer, a lightweight activation layer, a second linear mapping layer, and an output feature output unit.

6. The IoT intelligent road network safety monitoring system based on distributed optical fiber sensing according to claim 1, characterized in that: The LSTM-Transformer hybrid neural network model includes a feature fusion layer, a temporal coding layer, a global coding layer, a feature aggregation layer, and a classification output layer.

7. The IoT intelligent road network safety monitoring system based on distributed optical fiber sensing according to claim 1, characterized in that: The BP neural network includes a multidimensional feature input layer, a hidden layer, and a quantization result output layer.

8. The IoT intelligent road network safety monitoring system based on distributed optical fiber sensing according to claim 1, characterized in that: The disease development prediction module includes a Monte Carlo simulation unit. Monte Carlo simulation constructs a probability model, performs a large number of independent random samplings on the random variables in the model, substitutes the sampling results into the model for iterative calculations, and finally obtains the probability distribution and confidence interval of the results through statistical analysis.

Images