A smart tunnel construction safety monitoring and early warning system and method

Abstract

This invention discloses a smart tunnel construction safety monitoring and early warning system and method, belonging to the field of tunnel engineering monitoring technology. The method includes: collecting construction quality data and environmental data during tunnel construction; preprocessing the data and extracting key features from the preprocessed data to generate a daily data analysis report; generating dynamic thresholds for construction quality data based on the Dynamic Adaptive Threshold Optimization Model (DATOM); adjusting early warning triggering conditions according to real-time construction stage labels; and sending alarm information to terminal devices when construction quality data and environmental data deviate from the thresholds to reach the early warning conditions. This invention can meet the needs of dynamic changes in monitoring parameter thresholds at different construction stages, and has relatively high accuracy in safety monitoring and early warning.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering monitoring technology, specifically relating to a smart tunnel construction safety monitoring and early warning system and method. Background Technology

[0002] Traditional tunnel engineering monitoring has long relied on manual inspections and fixed threshold alarm mechanisms, which have inherent drawbacks such as slow response, data silos, and high false alarm rates. Although there are current monitoring solutions that use sensors and wireless transmission technologies, their technical bottlenecks still lie in the contradiction between static early warning thresholds and dynamic construction needs: existing systems use fixed thresholds, which cannot adapt to the quality control requirements of different stages such as tunnel excavation, support, and lining, and environmental interference (such as temperature and humidity fluctuations) can easily lead to misjudgments.

[0003] In recent years, the development of LSTM neural networks, Transformers, edge computing, and blockchain technologies has provided new solutions to the aforementioned problems. While LSTM can capture short-term temporal features, its ability to model long-distance dependencies is insufficient; while Transformer's self-attention mechanism excels at global correlation analysis, it has low sensitivity to local details. Therefore, effectively improving the ability to extract multi-scale features in complex construction environments such as tunnel construction, and meeting the need for dynamic threshold optimization of monitoring parameters at different construction stages, is of great significance for ensuring tunnel construction safety. Summary of the Invention

[0004] The purpose of this invention is to provide a smart tunnel construction safety monitoring and early warning system and method, thereby solving the problem that the existing technology cannot effectively extract multi-scale features in complex construction environments such as tunnel construction, especially the inability to meet the needs of dynamic changes in monitoring parameter thresholds at different construction stages, resulting in relatively low accuracy of tunnel construction safety monitoring.

[0005] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0006] The first aspect of this invention provides a smart tunnel construction safety monitoring and early warning method, comprising:

[0007] Collect construction quality data and environmental data during the tunnel construction process, and transmit the data to the server after converting the data format;

[0008] The data transmitted to the server is preprocessed, and key features are extracted from the preprocessed data for analysis to generate a daily data analysis report.

[0009] Thresholds for generating construction quality data based on the dynamic adaptive threshold optimization model DATOM;

[0010] The warning triggering conditions are adjusted according to the real-time construction stage labels. When the construction quality data and environmental data deviate from the threshold and reach the warning conditions, an alarm message is sent to the terminal device.

[0011] The dynamic threshold output by the dynamic adaptive threshold optimization model DATOM incorporates local features extracted by LSTM. and the feature vectors extracted by Transformer that capture global dependencies The specific formula for calculating the dynamic threshold is as follows:

[0012] ,

[0013] UB t , LB t These are the upper threshold and the lower threshold, respectively. This is the weight matrix. The bias vector is obtained through model training and optimization. for and The fused feature vector, This represents the historical benchmark average for this construction phase.

[0014] According to the technical solution of the first aspect of the present invention, the construction quality data includes structural mechanical parameters, specifically including one or more of the following: steel reinforcement stress, concrete strain, tunnel surrounding rock displacement, and crack propagation rate; the environmental data includes one or more of the following: ambient temperature, humidity, gas concentration, and groundwater level. The structural mechanical parameters serve as the primary monitoring indicators, while the environmental data serve as auxiliary monitoring indicators.

[0015] According to the technical solution of the first aspect of the present invention, the construction quality data and environmental data are transmitted to the server through an Internet of Things (IoT) wireless transmission device group. The IoT wireless transmission device group includes a 5G / 4G / WiFi DTU device and a built-in edge computing module. The edge computing module uses a Kalman filter algorithm to denoise the data and uses an isolated forest algorithm to detect outliers. The edge computing module compresses and encrypts the data before transmission and only transmits outlier data that exceeds the baseline by 20% to 30%.

[0016] According to the technical solution of the first aspect of the present invention, the data preprocessing includes one or more of data cleaning, format conversion, and outlier processing; the daily data analysis report includes the statistical distribution, trend prediction and comparative analysis with historical data of each monitoring data, and the construction quality data is superimposed on the tunnel BIM model through a three-dimensional heat map to mark high-risk areas in real time.

[0017] According to the technical solution of the first aspect of the present invention, the threshold for generating construction quality data based on the dynamic adaptive threshold optimization model DATOM specifically includes:

[0018] The LSTM module processes the local window data and outputs the hidden state feature vectors representing short-term mutations. ;

[0019] The Transformer module processes the global window data and outputs a multi-head attention feature vector that captures long-term dependencies. ;

[0020] The hidden state feature vector output by the LSTM module With the multi-head attention feature vector output by the Transformer module Weighted fusion is performed based on trainable fusion weight coefficients α to obtain fused features. : α∈[0,1];

[0021] Fusion features Input to the fully connected layer, through the weight matrix and bias vector Calculate the dynamic adjustment amount and overlay it with the historical benchmark average for this construction phase. This will output the final threshold pair.

[0022] According to the technical solution of the first aspect of the present invention, the generation process of the weighting coefficient α is as follows:

[0023] (1) Parameter initialization:

[0024] In the early stages of building the DATOM model, α is initialized as a trainable scalar variable with a neutral value of 0.5 and embedded in the neural network weight matrix;

[0025] (2) Dual-branch feature extraction and fusion:

[0026] When real-time sensor data streams are input, the model performs two types of feature analysis in parallel through LSTM and Transformer branches, and then fuses the two types of features to obtain fused features. ;

[0027] (3) Loss feedback and gradient generation:

[0028] After the fused features are processed by a fully connected layer to output a dynamic threshold prediction, this prediction is compared with the ideal threshold labeled by the engineer to calculate the Huber loss function; the gradient of α is automatically solved through backpropagation. When the LSTM feature's ability to reduce loss is better than that of the Transformer, the gradient is negative, driving α to increase;

[0029] (4) Dynamic optimization iteration:

[0030] The Adam optimizer is used to update the α value. The magnitude of each update is determined by the gradient direction and the historical update momentum. The Sigmoid function strictly constrains α to the [0,1] interval.

[0031] According to the technical solution of the first aspect of the present invention, the dynamic adaptive threshold optimization model DATOM is incrementally trained every 24 hours using newly added data, and during incremental training every 24 hours, the model inherits the α value converged on the previous day as the initial condition.

[0032] According to the technical solution described in the first aspect of the present invention, a graded risk warning is performed by combining the exceedance degree Ex of a single construction quality data point and the comprehensive risk index Rtotal. The warning rules are as follows:

[0033] Blue alert: The exceedance of a single construction quality parameter is 10% ≤ Ex < 20% or the crack propagation rate is 0.05mm / d ≤ v < 0.1mm / d;

[0034] Yellow alert: The exceedance of a single construction quality parameter is 20% ≤ Ex < 30%, or the exceedance of at least two construction quality parameters is 15% ≤ Ex < 20%, or the crack propagation rate is 0.1mm / d ≤ v < 0.15mm / d, and the comprehensive risk index is 20% ≤ Rtotal < 40%;

[0035] Orange alert: A single construction quality parameter exceeds the standard by 30% ≤ Ex < 50%, or the crack propagation rate is 0.15mm / d ≤ v < 0.2mm / d;

[0036] Red alert: A single construction quality parameter exceeds the standard by Ex ≥ 50%, or the crack propagation rate v ≥ 0.2 mm / d, or the comprehensive risk index Rtotal ≥ 50%;

[0037] The overall risk index Rtotal is calculated as follows:

[0038] ,

[0039] In the above formula:

[0040] The weighting of each construction stage is as follows: 0.6 for the excavation stage, 0.3 for the support stage, and 0.1 for the lining stage.

[0041] : No. i Overscaling of each parameter;

[0042] : No.i The critical overscaling threshold for each parameter;

[0043] : Environmental correction factor, obtained by customization.

[0044] According to the technical solution of the first aspect of the present invention, when the dynamic threshold generated by the DATOM model triggers an early warning three times consecutively, the system automatically upgrades the early warning level, generates an emergency task work order, and assigns it to designated personnel.

[0045] A second aspect of the present invention also provides a smart tunnel construction safety monitoring and early warning system, comprising:

[0046] The data acquisition unit is used to collect construction quality data and environmental data during the tunnel construction process, and transmit the data to the server after converting the data format.

[0047] The data preprocessing unit is used to preprocess the data transmitted to the server, extract key features from the preprocessed data, and generate a daily data analysis report.

[0048] The dynamic threshold generation unit generates dynamic thresholds for construction quality data based on the DATOM dynamic adaptive threshold optimization model; and

[0049] The early warning unit sends an alarm message to the terminal device when construction quality data and environmental data deviate from the threshold and reach the early warning condition.

[0050] The dynamic adaptive threshold optimization model DATOM includes an LSTM module, a Transformer module, a feature fusion module, and a fully connected layer. The LSTM module is used to extract local features. The Transformer module is used to extract feature vectors that capture global dependencies. The feature fusion module is used for... and Perform weighted fusion to generate a fused feature vector. Fully connected layers are based on fused feature vectors. Through the weight matrix and bias vector Calculate the dynamic adjustment amount and overlay it with the historical benchmark average for this construction phase. This will output the final threshold pair.

[0051] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0052] (1) This invention collects construction quality parameters and environmental data in real time during tunnel construction, uses construction quality parameters as the main monitoring indicators, and generates dynamic early warning thresholds for construction quality parameters based on the dynamic adaptive threshold optimization model DATOM that integrates LSTM and Transformer. This not only effectively extracts multi-scale features in complex construction environments such as tunnel construction, but also effectively improves the reliability and accuracy of risk warning.

[0053] (2) When generating dynamic early warning thresholds for construction quality data through the dynamic adaptive threshold optimization model DATOM, this invention performs weighted fusion of local features (such as short-term stress fluctuations) extracted by LSTM and global dependencies (such as cross-stage stress trends) captured by Transformer. Furthermore, during model training, the value of the fusion weight α is backpropagated and updated, thereby enabling the model to automatically optimize the fusion ratio of local and global features according to the characteristics of the data. This is beneficial to further ensure the rationality of the early warning threshold and thus improve the accuracy of subsequent risk warnings.

[0054] (3) This invention combines the exceedance of single construction quality data Ex and the comprehensive risk index Rtotal to carry out graded risk warning, which can effectively avoid the occurrence of risk omission and false detection; at the same time, it can further generate different warning thresholds according to different construction stages, thus meeting the needs of dynamic changes in monitoring parameter thresholds at different construction stages and having good dynamic adaptability.

[0055] (4) This invention overlays construction quality data such as stress and strain onto the tunnel BIM model through a three-dimensional heat map, and marks the spatial distribution of high-risk areas in real time. When the dynamic threshold generated by the DATOM model triggers the warning three times in a row, the system automatically raises the warning level, generates an emergency task work order and assigns it to the designated personnel, thereby realizing closed-loop management of "risk identification-level determination-task assignment". Attached Figure Description

[0056] The dimensions and scales in the accompanying drawings do not represent the actual dimensions and scales of the product. The drawings are for illustrative purposes only, and some non-essential elements or features have been omitted for clarity.

[0057] Figure 1 This is a flowchart of the early warning method of the present invention;

[0058] Figure 2 This is a schematic diagram of the structure of the LSTM module and the Transformer module in an embodiment of the present invention. Detailed Implementation

[0059] The present invention will now be described in detail with reference to specific embodiments.

[0060] Combination Figure 1 As shown in the figure, this embodiment of the invention provides a smart tunnel construction safety monitoring and early warning method, including the following steps:

[0061] Step 1: Collect construction quality data and environmental data in real time during the tunnel construction process using sensors, and then transmit the data to the server after converting the data format.

[0062] In this invention, construction quality data serves as the primary safety early warning and monitoring data, while environmental data serves as auxiliary monitoring data. Specifically, in this embodiment of the invention, the construction quality data includes structural mechanical parameters, which specifically include one or more of the following: steel reinforcement stress, concrete strain data, tunnel surrounding rock displacement, and crack propagation rate data. The environmental data includes, but is not limited to, ambient temperature, humidity, gas concentration, and groundwater level.

[0063] Specifically, the aforementioned steel reinforcement stress, concrete strain, and tunnel surrounding rock displacement are obtained through stress sensors, strain sensors, and displacement sensors, respectively. The crack propagation rate is based on the strain changes on both sides of the crack monitored by strain gauges, combined with time data. (In actual engineering, strain gauges can be installed in potential cracking zones and existing apparent cracks immediately after tunnel excavation, while simultaneously deploying stress and strain sensors. When the crack propagates, the resistance of the strain gauges changes, and the crack width change can be calculated by measuring the resistance change.) The ambient temperature, humidity, gas concentration, and groundwater level are collected through temperature sensors, humidity sensors, gas concentration sensors, and liquid level sensors, respectively.

[0064] It should be noted that the specific types of sensors used in this invention can be selected and combined according to actual needs. For example, they can also be used to collect data on the vibration of the support structure, and the data collection interval for each project can be dynamically adjusted according to the construction stage. For example, data can be collected every 5 minutes during the excavation stage, increased to every 2 minutes during the support stage, and returned to every 5 minutes during the lining stage.

[0065] Preferably, the data collected by the sensors is transmitted to the server via an IoT wireless transmission device group, which includes a 5G / 4G / WiFi DTU device and a built-in edge computing module. The edge computing module uses a Kalman filter algorithm for data denoising and an isolated forest algorithm to detect outliers.

[0066] Specifically, in this embodiment of the invention, when using the Kalman filter algorithm for data denoising, the state equation describes the relationship between the current state and the previous state, and the observation equation correlates the actual measured value with the predicted value; both process noise and observation noise are assumed to be Gaussian distributed, and the initial covariance matrix is ​​set as a diagonal matrix to simplify calculations. ， Denoise the raw sensor data.

[0067] More preferably, the edge computing module compresses and encrypts the data (e.g., using the AES-256 algorithm) before transmitting it to the cloud, and only transmits abnormal data that exceeds the baseline by 20% to 30%, thereby saving bandwidth and solving the problem of bandwidth pressure surge caused by full data transmission from sensors.

[0068] Step 2: Perform data preprocessing on the data transmitted to the server, and extract key feature analysis from the preprocessed construction quality data to generate a daily data analysis report.

[0069] The preprocessing includes one or more of the following: data cleaning, format conversion, and outlier handling. The data cleaning uses a sliding window algorithm (window length 30 minutes, step size 5 minutes) to detect outliers, and combines the construction stage labels to correct sensor drift errors caused by temperature.

[0070] The process involves extracting key features (mean, variance, extreme values) from the preprocessed data to generate a daily data analysis report. This report includes the statistical distribution of the monitoring data, trend prediction (using the ARIMA model), and comparative analysis with historical data. Furthermore, the construction quality data is displayed in multiple dimensions using graphs, tables, and heat maps to achieve real-time monitoring and collection. The corresponding data is then overlaid onto the tunnel BIM model using a 3D heat map to mark high-risk areas in real time.

[0071] Step 3: Generate thresholds for construction quality data based on the DATOM dynamic adaptive threshold optimization model;

[0072] In this embodiment of the invention, the Dynamic Adaptive Threshold Optimization Model (DATOM) is constructed based on the fusion of LSTM and Transformer. By combining the advantages of LSTM and Transformer, it can effectively capture multi-scale features of time series data, thereby facilitating the accurate generation of dynamic thresholds. LSTM, with its gating mechanism, excels at capturing local short-term dependencies in data, while Transformer's self-attention mechanism can directly obtain the relationship between any two time points in the sequence, excelling at capturing global dependencies. Combining the two allows for a comprehensive understanding of the characteristics of time series data, providing a more accurate basis for generating dynamic thresholds.

[0073] Specifically, the construction process of the Dynamic Adaptive Threshold Optimization Model (DATOM) includes:

[0074] (1) Construction of LSTM module

[0075] The core of LSTM is the gating mechanism, which includes input gates, forget gates, and output gates to control the information flow and effectively alleviate the gradient vanishing problem.

[0076] LSTM modules use local window data As input data, at any time t The input gate calculation formula is: The Gate of Oblivion The output gate is Candidate memories are Update memory units to The hidden state is updated to .in, This represents the sigmoid activation function. This represents element-wise multiplication, where W is the weight matrix and b is the bias term.

[0077] (2) Construction of the Transformer module

[0078] The Transformer mainly consists of self-attention mechanism, multi-head attention, feedforward neural network, residual connection and layer normalization.

[0079] The Transformer module uses global window data. As input data, in the self-attention mechanism, a Query, Key, and Value matrix is ​​first generated through a linear transformation. ,in , , This is the learnable weight matrix. Next, the attention score is calculated. , d k It is the dimension of the key.

[0080] Multi-head attention maps the input to multiple subspaces, computes the attention for each subspace, and then concatenates the results. The computation for each head is as follows: Multi-head attention output is ,in The first i The parameters of the head and the output projection matrix.

[0081] In addition, the Transformer module incorporates positional encoding to preserve sequence position information and improves training stability through feedforward networks, residual connections, and layer normalization.

[0082] (3) Construction of the feature fusion module

[0083] The feature fusion module is used to process the local features generated by the LSTM module. With the output of the Transformer module Unlike the serial fusion method, the preferred embodiment of this invention uses a weighted fusion method, where the fusion output is... α ∈ [0,1], where α is a trainable fusion weight that adaptively balances local and global features. During model training, the value of α is adjusted through backpropagation, enabling the model to automatically optimize the fusion ratio of local and global features based on data characteristics.

[0084] (4) Model training and dynamic threshold generation

[0085] Training data: The training data for the DATOM model includes historical monitoring data of construction quality and construction stage labels S. t (e.g., excavation / support / lining), environmental parameter E t (Such as temperature / humidity) and manually labeled early warning feedback results. These data provide rich information for the model to learn the characteristics and patterns of time series under different conditions.

[0086] Incremental training: The model is incrementally trained every 24 hours using newly added data. During training, the LSTM module continues to learn local feature changes in the new data, while the Transformer module further captures global dependencies in the data. By continuously updating the model parameters, the model can adapt to dynamic changes in the data.

[0087] Dynamic threshold generation: After training, the DATOM model generates a dynamic threshold curve (upper threshold UB(t) and lower threshold LB(t)) for each sensor.

[0088] Combination Figure 2 As shown, the process of generating dynamic thresholds using the DATOM model is as follows:

[0089] First, the real-time sensor data stream is acquired: specifically, in this embodiment of the invention, a local window of length 10 is extracted. and global window data with a length of 100 At the same time, the current construction stage label St and environmental parameter Et are loaded.

[0090] Then, feature extraction is performed in parallel: the LSTM module uses a gating mechanism (including the input gate) to perform feature extraction. i t Forgotten Gate f t Memory unit Ct Process local data and output feature vectors representing short-term mutations. The Transformer module analyzes global data through a self-attention mechanism (including the query matrix Q, key matrix K, and value matrix V) to generate feature vectors that capture long-term dependencies. .

[0091] The two types of features enter the adaptive fusion layer, where they are weighted and synthesized using the training-obtained weight coefficients α to obtain the fused features. : ;

[0092] Fusion features Input to the fully connected layer, through the weight matrix and bias vector Calculate the dynamic adjustment amount and overlay it with the historical benchmark average for this construction phase. The final output threshold pair is:

[0093] ;

[0094] Among them, the weight matrix and bias vector Updated by Adam optimizer; This represents the historical benchmark average for this construction phase.

[0095] A further preferred method is to use an LSTM neural network to extract temporal features, incrementally update the network weights every 24 hours, and dynamically adjust the threshold curve so that the threshold deviation tolerance decreases linearly with the construction progress.

[0096] Furthermore, as a preferred embodiment, the process for generating α is as follows:

[0097] (1) Parameter initialization:

[0098] In the early stages of model building, α, as a trainable scalar variable, is initialized to a neutral value of 0.5 and embedded in the neural network weight matrix. At this point, the system has not yet learned any data patterns, and the LSTM and Transformer features are given equal weights.

[0099] (2) Dual-branch feature extraction:

[0100] When the sensor data stream is input in real time, the system performs two types of feature analysis in parallel:

[0101] LSTM branches focus on local windows, such as a local window of 10 time steps (t1=10), and capture transient anomalies through a gating mechanism;

[0102] Transformer branch analysis global window, such as a global window of 100 time steps (t2=100), uses self-attention mechanism to model long-term patterns.

[0103] (3) Differentiable fusion operation:

[0104] The system performs core fusion computing: Here, α acts as an adjustment lever. If its value is 0.7, then the LSTM features are given 70% weight, while the Transformer features only account for 30%.

[0105] (4) Loss feedback and gradient generation:

[0106] After the fused features are processed by a fully connected layer to output a dynamic threshold prediction, the system compares this prediction with the ideal threshold labeled by the engineer and calculates the Huber loss function. The gradient of α is automatically solved through backpropagation. The gradient directly reflects the contribution efficiency of the two types of features: when the LSTM features are better at reducing loss than the Transformer, the gradient is negative, driving α to increase.

[0107] (5) Dynamic optimization iteration:

[0108] The Adam optimizer is used to update the α value, and the magnitude of each update is determined by both the gradient direction and the historical update momentum. This process is continuously repeated during training, allowing α to gradually evolve from the initial 0.5 to a scene-fitting state.

[0109] (6) Engineering constraint guarantees:

[0110] The sigmoid function is used to strictly constrain α within the [0,1] interval to prevent weight runaway. During incremental training every 24 hours, the system inherits the α value converged the previous day as the initial condition, and new data only causes limited drift.

[0111] Ultimately, α converges into a data-driven intelligent weight through several gradient updates, its value quantifying the relative importance of local sudden risks and global evolutionary patterns in real time. By optimizing the generation process of α, the different requirements for construction quality thresholds at different stages can be met, thus improving the accuracy of early warning.

[0112] Step 4: Adjust the early warning trigger conditions according to the real-time construction stage labels. When the construction quality data and environmental data deviate from the threshold and reach the early warning conditions, send alarm information to the terminal equipment.

[0113] Specifically, as one implementation method, an early warning is activated when construction quality data or environmental data deviates from the threshold to the set excess level. The threshold for environmental data is directly set in advance by humans, and the threshold for environmental data can be adjusted adaptively at different construction stages. For example, the gas concentration threshold can be appropriately increased during the excavation stage.

[0114] The scaling factor is defined as follows, with two cases: upper scaling factor and lower scaling factor:

[0115] .

[0116] As a further preferred implementation method, the exceedance of single construction quality data is considered. Ex A comprehensive risk index, Rtotal, is introduced for tiered risk warnings. The specific warning rules are as follows:

[0117] Blue alert: The exceedance of a single construction quality parameter is 10% ≤ Ex < 20% or the crack propagation rate is 0.05mm / d ≤ v < 0.1mm / d;

[0118] Yellow alert: The exceedance of a single construction quality parameter is 20% ≤ Ex < 30%, or the exceedance of at least two construction quality parameters is 15% ≤ Ex < 20%, or the crack propagation rate is 0.1mm / d ≤ v < 0.15mm / d, and the comprehensive risk index is 20% ≤ Rtotal < 40%;

[0119] Orange alert: The exceedance of a single construction quality parameter is 30% ≤ Ex < 50%, or the crack propagation rate is 0.15 mm / d ≤ v < 0.2 mm / d;

[0120] Red alert: Exceeding the standard for a single construction quality parameter Ex ≥ 50%, or crack propagation rate v ≥ 0.2 mm / d, or comprehensive risk index Rtotal ≥ 50%;

[0121] The overall risk index Rtotal is calculated as follows:

[0122] ,

[0123] In the above formula:

[0124] The weighting of each construction stage is as follows: 0.6 for the excavation stage, 0.3 for the support stage, and 0.1 for the lining stage.

[0125] : No. i Overscaling of each parameter;

[0126] : No. i The critical overscaling thresholds for each parameter are set in advance;

[0127] Environmental correction factor, which is customized. For example, rainy season = 1.2, fire alarm triggered = 2.0, normal = 1.0.

[0128] Since relying solely on out-of-scale indicators for early warning under the influence of multiple factors can lead to missed or false detections, this embodiment of the invention further introduces a comprehensive risk index, Rtotal, which effectively integrates the out-of-scale levels of construction quality parameters and environmental parameters, thereby helping to reduce the occurrence of false or missed detections.

[0129] Furthermore, coupled discrimination rules for natural disasters can be introduced:

[0130] (1) Fire risk index:

[0131] ,

[0132] in, T 0 represents the safe temperature threshold. CO 0 represents the safe threshold for CO concentration.

[0133] (2) Water inrush risk index:

[0134] ,

[0135] Where h is the groundwater level, h0 is the safe water level, Q is the inflow rate, and Q0 is the safe inflow rate.

[0136] A further preferred embodiment is that when the dynamic threshold generated by the DATOM model triggers an early warning three times consecutively, the system automatically upgrades the early warning level, generates an emergency task work order, and assigns it to designated personnel.

[0137] As a further preferred implementation scheme, in step three, when an early warning occurs, the early warning linkage mechanism is triggered:

[0138] The system automatically retrieves the global feature weight matrix output by the Transformer to locate the parameters that contribute the most to the risk (such as stress and vibration). (weight increased significantly)

[0139] By combining BIM model 3D positioning, areas with abnormal related parameters are highlighted in the tunnel heat map, thereby achieving accurate positioning of risk areas;

[0140] Emergency work orders containing parameter correlation analysis are generated (such as "Stress-vibration coupling anomaly in section K10+500, it is recommended to check the steel arch connection immediately") and pushed out simultaneously via SMS, email and platform.

[0141] Furthermore, in this embodiment of the invention, alarm information is stored using blockchain technology. The blockchain adopts a consortium blockchain architecture, deploys 3 consensus nodes, and the PBFT mechanism requires more than 2 / 3 of the nodes to reach consensus. Data block hash verification uses SHA-256, the block generation interval is 10 minutes, and logs are stored using AES-256 encryption, thereby ensuring the immutability of the warning data and forming a complete closed-loop system of intelligent monitoring, risk assessment, and emergency response.

[0142] This invention also provides a smart tunnel construction safety monitoring and early warning system, comprising:

[0143] The data acquisition unit is used to collect construction quality data, environmental data, and equipment status parameters during tunnel construction, and then transmit the data to the server after converting the data format.

[0144] The data preprocessing unit is used to preprocess the data transmitted to the server, extract key features from the preprocessed data, and generate a daily data analysis report.

[0145] The dynamic threshold generation unit generates dynamic thresholds for construction quality data based on the DATOM dynamic adaptive threshold optimization model; and

[0146] The early warning unit sends an alarm message to the terminal device when construction quality data and environmental data deviate from the threshold and reach the early warning condition.

[0147] The dynamic adaptive threshold optimization model DATOM includes an LSTM module, a Transformer module, a feature fusion module, and a fully connected layer. The LSTM module is used to extract local features. The Transformer module is used to extract feature vectors that capture global dependencies. The feature fusion module is used for... and Perform weighted fusion to generate a fused feature vector. Fully connected layers are based on fused feature vectors. Through the weight matrix and bias vector Calculate the dynamic adjustment amount and overlay it with the historical benchmark average for this construction phase. This will output the final threshold pair.

[0148] This system can monitor environmental parameters and equipment status inside the tunnel in real time. Through intelligent data processing and analysis, it can promptly identify problems and issue early warnings. It can also provide customized content based on the actual situation of tunnel construction to ensure the safe and efficient operation of the tunnel.

[0149] Furthermore, the early warning system is linked to the status of each sensor. If a sensor is operating normally, it will display "online"; if the device is disconnected, it will display "offline". If the offline status is exceeded by 30 minutes, an equipment maintenance alarm will be triggered.

[0150] The scope of protection of this invention is defined only by the claims. Thanks to the teachings of this invention, those skilled in the art will readily recognize that alternative structures to the structures disclosed herein can be used as feasible alternative implementations, and that the implementations disclosed herein can be combined to produce new implementations, which also fall within the scope of the appended claims.

Claims

1. A method for monitoring and early warning of tunnel construction safety, characterized in that, include: Collect construction quality data and environmental data during the tunnel construction process, and transmit the data to the server after converting the data format; The data transmitted to the server is preprocessed, and key features are extracted from the preprocessed data for analysis to generate a daily data analysis report. Thresholds for generating construction quality data based on the dynamic adaptive threshold optimization model DATOM; The warning triggering conditions are adjusted according to the real-time construction stage labels. When the construction quality data and environmental data deviate from the set threshold and reach the warning conditions, an alarm message is sent to the terminal device. The dynamic threshold output by the dynamic adaptive threshold optimization model DATOM fuses local features extracted by the LSTM and a feature vector capturing global dependencies extracted by the Transformer The dynamic threshold calculation formula is as follows: ， wherein, UB t , LB t are an upper threshold and a lower threshold, respectively, is a weight matrix, is a bias vector, which is obtained by model training and optimization; is a fusion feature vector of and is a historical reference mean value of the construction stage.​ 2. The intelligent tunnel construction safety monitoring and early warning method according to claim 1, characterized in that, The construction quality data includes structural mechanical parameters, which include one or more of the following: steel reinforcement stress, concrete strain, tunnel surrounding rock displacement, and crack propagation rate; the environmental data includes one or more of the following: ambient temperature, humidity, gas concentration, and groundwater level.

3. The intelligent tunnel construction safety monitoring and early warning method according to claim 1, characterized in that, The construction quality data and environmental data are transmitted to the server through an IoT wireless transmission device group. The IoT wireless transmission device group includes a 5G / 4G / WiFi DTU device and a built-in edge computing module. The edge computing module uses the Kalman filter algorithm to denoise the data and the isolated forest algorithm to detect outliers. The edge computing module compresses and encrypts the data before transmission and only transmits abnormal data that exceeds the baseline by 20% to 30%.

4. The intelligent tunnel construction safety monitoring and early warning method according to claim 1, characterized in that, The data preprocessing includes one or more of the following: data cleaning, format conversion, and outlier handling; the daily data analysis report includes the statistical distribution, trend prediction, and comparative analysis with historical data of each monitoring data, and the construction quality data is overlaid onto the tunnel BIM model through a three-dimensional heat map to mark high-risk areas in real time.

5. The intelligent tunnel construction safety monitoring and early warning method according to any one of claims 1-4, characterized in that, The thresholds for generating construction quality data based on the DATOM dynamic adaptive threshold optimization model specifically include: The LSTM module processes the local window data and outputs the hidden state feature vectors representing short-term mutations. ; The Transformer module processes the global window data and outputs a multi-head attention feature vector that captures long-term dependencies. ; The hidden state feature vector output by the LSTM module With the multi-head attention feature vector output by the Transformer module Weighted fusion is performed based on trainable fusion weight coefficients α to obtain fused features. : α∈[0,1]; Fusion features Input to the fully connected layer, through the weight matrix and bias vector Calculate the dynamic adjustment amount and overlay it with the historical benchmark average for this construction phase. This will output the final threshold pair.

6. The intelligent tunnel construction safety monitoring and early warning method according to claim 5, characterized in that, The process of generating the weighting coefficient α is as follows: (1) Parameter initialization: In the early stages of building the DATOM model, α is initialized as a trainable scalar variable with a neutral value of 0.5 and embedded in the neural network weight matrix; (2) Dual-branch feature extraction and fusion: When real-time sensor data streams are input, the model performs two types of feature analysis in parallel through LSTM and Transformer branches, and then fuses the two types of features to obtain fused features. ; (3) Loss feedback and gradient generation: After the fused features are processed by a fully connected layer to output a dynamic threshold prediction, this prediction is compared with the ideal threshold labeled by the engineer to calculate the Huber loss function; the gradient of α is automatically solved through backpropagation. When the LSTM feature's ability to reduce loss is better than that of the Transformer, the gradient is negative, driving α to increase; (4) Dynamic optimization iteration: The Adam optimizer is used to update the α value. The magnitude of each update is determined by the gradient direction and the historical update momentum. The Sigmoid function strictly constrains α to the [0,1] interval.

7. The intelligent tunnel construction safety monitoring and early warning method according to claim 6, characterized in that, The dynamic adaptive threshold optimization model DATOM is incrementally trained every 24 hours using newly added data, and during each 24-hour incremental training, the model inherits the α value that converged the previous day as the initial condition.

8. The intelligent tunnel construction safety monitoring and early warning method according to claim 7, characterized in that, A graded risk warning system is implemented by combining the exceedance degree (Ex) of individual construction quality data with the comprehensive risk index (Rtotal). The warning rules are as follows: Blue alert: The exceedance of a single construction quality parameter is 10% ≤ Ex < 20% or the crack propagation rate is 0.05mm / d ≤ v < 0.1mm / d; Yellow alert: The exceedance of a single construction quality parameter is 20% ≤ Ex < 30%, or the exceedance of at least two construction quality parameters is 15% ≤ Ex < 20%, or the crack propagation rate is 0.1mm / d ≤ v < 0.15mm / d, and the comprehensive risk index is 20% ≤ Rtotal < 40%; Orange alert: A single construction quality parameter exceeds the standard by 30% ≤ Ex < 50%, or the crack propagation rate is 0.15mm / d ≤ v < 0.2mm / d; Red alert: A single construction quality parameter exceeds the standard by Ex ≥ 50%, or the crack propagation rate v ≥ 0.2 mm / d, or the comprehensive risk index Rtotal ≥ 50%; The overall risk index Rtotal is calculated as follows: ， In the above formula: The weighting of each construction stage is as follows: 0.6 for the excavation stage, 0.3 for the support stage, and 0.1 for the lining stage. : No. i Overscaling of each parameter; : No. i The critical overscaling threshold for each parameter; : Environmental correction factor, obtained by customization.

9. A smart tunnel construction safety monitoring and early warning system, characterized in that, include: The data acquisition unit is used to collect construction quality data, environmental data, and equipment status parameters during tunnel construction, and then transmit the data to the server after converting the data format. The data preprocessing unit is used to preprocess the data transmitted to the server, extract key features from the preprocessed data, and generate a daily data analysis report. The dynamic threshold generation unit generates dynamic thresholds for construction quality data based on the dynamic adaptive threshold optimization model DATOM. as well as The early warning unit sends an alarm message to the terminal device when construction quality data and environmental data deviate from the threshold and reach the early warning condition. The dynamic adaptive threshold optimization model DATOM includes an LSTM module, a Transformer module, a feature fusion module, and a fully connected layer. The LSTM module is used to extract local features. The Transformer module is used to extract feature vectors that capture global dependencies. The feature fusion module is used for... and Perform weighted fusion to generate a fused feature vector. Fully connected layers are based on fused feature vectors. Through the weight matrix and bias vector Calculate the dynamic adjustment amount and overlay it with the historical benchmark average for this construction phase. This will output the final threshold pair.

10. The intelligent tunnel construction safety monitoring and early warning system according to claim 9, characterized in that, The dynamic threshold output by the DATOM dynamic adaptive threshold optimization model incorporates local features extracted by LSTM. and the feature vectors extracted by Transformer that capture global dependencies The specific formula for calculating the dynamic threshold is as follows: ， in, UB t , LB t These are the upper threshold and the lower threshold, respectively. This is the weight matrix. The bias vector is obtained through model training and optimization. for and The fused feature vector, This represents the historical benchmark average for this construction phase.

Images