Tunnel safety assessment method and system based on deep learning technology

Abstract

The invention discloses a tunnel safety assessment method based on a deep learning technology, and the method comprises the steps: collecting safety monitoring data in a tunnel, and carrying out the preprocessing of the collected safety monitoring data; outputting time sequence characteristics based on a long short-term memory (LSTM) network and / or a Transform model of a converter; local space feature extraction is carried out based on a convolutional neural network CNN; outputting global structure features based on a graph neural network GNN; performing joint feature extraction on the time sequence features, the local spatial features and the global structure features based on a multi-input deep neural network DNN and / or a gated loop unit GRU model, and outputting joint features; and classifying the joint features based on a support vector machine (SVM) and / or the DNN, and outputting a safety assessment level of the tunnel. Through effective fusion of multiple deep learning models, multi-dimensional data features can be comprehensively considered, and the accuracy and real-time performance of tunnel safety assessment are improved.

Description

Technical Field

[0001] The present invention relates to a tunnel safety assessment method and system based on deep learning technology. Background Art

[0002] Traditional tunnel safety assessment methods rely on manual inspections and regular analysis of monitoring data, primarily including structural inspections and sensor data recording. However, these methods suffer from cumbersome manual analysis, an inability to promptly reflect tunnel status changes, and inaccurate assessment results, making them ineffective in meeting the demands of modern tunnel safety management.

[0003] With the rapid development of artificial intelligence and deep learning technologies, tunnel safety assessment methods based on big data analysis have gradually become a research hotspot. Chinese invention patent CN106910002A discloses a "Method for Structural Safety Assessment of Subway Tunnel Structures." Based on a fuzzy comprehensive evaluation method, this method considers the correlation of various evaluation indicators and the impact of their levels, and performs variable weighting. However, this method fails to account for temporal, spatial, and structural characteristics, limiting the accuracy and reliability of the assessment results. Summary of the Invention

[0004] The present invention provides a tunnel safety assessment method and system based on deep learning technology. Through the effective integration of multiple deep learning models, it can comprehensively consider multi-dimensional data characteristics and improve the accuracy and real-time performance of tunnel safety assessment.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] In a first aspect, the present invention provides a tunnel safety assessment method based on deep learning technology, comprising:

[0007] Collecting safety monitoring data in the tunnel and preprocessing the collected safety monitoring data; the safety monitoring data includes time series data and spatial data, and the preprocessing includes: data cleaning, data denoising, data standardization, and / or data transformation and enhancement;

[0008] Performing sequence modeling on the preprocessed time series data based on a long short-term memory network (LSTM) and / or a transformer model to output time series features for representing time dependency and dynamic change trends; extracting local spatial features from the preprocessed spatial data based on a convolutional neural network (CNN); and performing association modeling on the positions and spatial relationships of the sensors in the tunnel used to collect the safety monitoring data based on a graph neural network (GNN) to output global structural features;

[0009] Based on a multi-input deep neural network DNN and / or a gated recurrent unit GRU model, joint feature extraction is performed on the temporal features, the local spatial features, and the global structural features, and the joint features are output; based on a support vector machine SVM and / or the DNN, the joint features are classified, and the safety assessment level of the tunnel is output.

[0010] In an optional embodiment, performing sequence modeling on the preprocessed time series data based on LSTM includes:

[0011] The input of the LSTM is the time series data X t ={x1,x2,…,x t}, and output the hidden state H based on the following formula LSTM ={h1,h2,…,h t}:

[0012] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0013] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0014]

[0015] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0016] h t =o t ☉tanh(C t )

[0017] Among them, x t is the input of time series data at time t, f t is the output of the forget gate, i t is the output of the input gate, is the candidate cell state, C t To update the cell state, o t is the output of the output gate, h t is the hidden state at time t, σ is the sigmoid activation function, ⊙ represents element-wise multiplication, W f ,W i ,W C,W o is the weight matrix, f b ,b i ,b C ,b o is the bias term.

[0018] In an optional embodiment, performing sequence modeling on the preprocessed time series data based on a Transformer model includes:

[0019] The input of the Transformer model is the time series data X t ={x1,x2,…,x t}, and output the representation Y after self-attention mechanism processing based on the following self-attention mechanism formula t ={y1,y2,…,y t}:

[0020] Q t =W Q ·X t , K t =W K ·X t , V t =W V ·X t

[0021]

[0022] Y t =Attention(Q t ,K t ,V t )

[0023] Among them, Q t is the query matrix, K t is the bond matrix, V t is the value matrix, W Q is the weight matrix corresponding to the query matrix, W K is the weight matrix corresponding to the key matrix, W V is the weight matrix corresponding to the value matrix, d k is the dimension of the key vector, and softmax(·) is the normalization operation.

[0024] In an optional embodiment, performing local spatial feature extraction on the pre-processed spatial data based on a convolutional neural network (CNN) includes:

[0025] The input of the convolution operation of the CNN is the spatial data X i ={x1,x2,…,x i}, and output the local spatial feature Y based on the calculation formula of the following convolution operation i ={y1,y2,…,y i}:

[0026]

[0027] Among them, j represents the neighbor node of node i, k is the total number of neighbor nodes of each node i, X i+j is the local area of ​​the input spatial data, W j is the convolution kernel, b is the bias term, and f(·) is the activation function.

[0028] In an optional embodiment, the graph neural network (GNN)-based modeling performs correlation modeling on the positions and spatial relationships of the sensors used to collect the safety monitoring data in the tunnel, and outputs global structural features, including:

[0029] The input of the GNN is the location node feature H = {h1,h2,…,h n}, and output the global structural feature H'={h'1,h'2,…,h' based on the calculation formula of the following graph convolution operation n}:

[0030]

[0031] Among them, h i is the feature of the i-th node, N(i) is the set of neighbor nodes of the i-th node, c ij is the normalization factor, σ is the sigmoid activation function, W is the weight matrix, and b is the bias term.

[0032] In an optional embodiment, the performing joint feature extraction on the temporal features, the local spatial features, and the global structural features based on a multi-input deep neural network DNN and outputting the joint features includes:

[0033] Based on the following formula, the timing feature f time , the local spatial feature f local , and the global structural feature f global Splice and get the overall input vector x fusion :

[0034] x fusion =concat(f time , f local , f global )

[0035] The DNN includes L fully connected layers, and the output of the lth layer is expressed by the following formula:

[0036] h (l) =σ(W (l) h (l-1) +b (l) ), l=, ..., L

[0037] Among them, h (0) =x fusion , W (l) is the weight matrix, b (l) is the bias term, σ is the sigmoid activation function, and the joint feature is f joint =h (L) .

[0038] In an optional embodiment, the gated recurrent unit (GRU) model is used to perform joint feature extraction on the temporal features, the local spatial features, and the global structural features, and output a fused feature vector, including:

[0039] Based on the following formula, the timing feature f time , the local spatial feature f local , and the global structural feature f global Splice and get the overall input vector x fusion :

[0040] x fusion =concat(f time , f local , f global )

[0041] The reset gate r of the GRU t , update gate z t , candidate hidden states The final hidden state h t They are expressed by the following formulas:

[0042] r t =σ(W r ·[h t-1 ,x t ]+b r )

[0043] z t =σ(W z ·[h t-1 ,x t ]+b z )

[0044]

[0045] Among them, W r ,W z ,W h is the weight matrix, br ,b z ,b h is the bias term, σ is the sigmoid activation function, and the joint feature is f joint =h t .

[0046] In an optional embodiment, the classifying the joint features based on a support vector machine (SVM) and outputting the security assessment level of the tunnel includes:

[0047] The input of the SVM is the joint feature f joint , and based on the following classification decision function f(f joint ) Output the security assessment level of the tunnel:

[0048] f(f joint )=w T f joint +b

[0049] Among them, w is the weight matrix, w T is the transpose of vector w, and b is the bias term.

[0050] In an optional embodiment, the method further includes:

[0051] An incremental training set is updated according to the security monitoring data and the security assessment level, and incremental training is performed on the LSTM, the Transformer model, the CNN, the GNN, the DNN, the GRU, and / or the SVM based on the incremental training set and a preset reinforcement learning strategy.

[0052] In a second aspect, the present invention provides a tunnel safety assessment system based on deep learning technology, characterized in that it includes: a data processing module, a feature processing module, and a level assessment module;

[0053] The data processing module is used to collect safety monitoring data in the tunnel and pre-process the collected safety monitoring data; the safety monitoring data includes time series data and spatial data, and the pre-processing includes: data cleaning, data denoising, data standardization, and / or data transformation and enhancement;

[0054] The feature processing module is used to perform sequence modeling on the pre-processed time series data based on a long short-term memory network (LSTM) and / or a transformer model, and output time series features for representing time dependency and dynamic change trends; extract local spatial features from the pre-processed spatial data based on a convolutional neural network (CNN); and perform association modeling on the positions and spatial relationships of the sensors in the tunnel used to collect the safety monitoring data based on a graph neural network (GNN), and output global structural features;

[0055] The grade assessment module is used to perform joint feature extraction on the temporal features, the local spatial features and the global structural features based on a multi-input deep neural network DNN and / or a gated recurrent unit GRU model, and output the joint features; classify the joint features based on a support vector machine SVM and / or the DNN, and output the security assessment grade of the tunnel.

[0056] The beneficial effects brought about by the technical solution provided by the present invention include at least:

[0057] 1. The present invention integrates multi-dimensional information such as time series data, spatial data, and the spatial position relationship of sensors into a comprehensive model, integrating multi-source data to comprehensively and accurately reflect the safety status of the tunnel, avoiding assessment errors caused by a single data type in traditional methods.

[0058] 2. The present invention achieves efficient dynamic modeling and prediction capabilities by modeling time series data through the long short-term memory network LSTM and the transformer model, which can effectively capture the dynamic change trend of the tunnel structure, thereby improving the ability to predict future risks.

[0059] 3. The present invention uses convolutional neural network (CNN) to extract features from spatial data, achieving accurate spatial feature extraction, automatically identifying possible local structural problems in the tunnel, and improving the accuracy of spatial information processing.

[0060] 4. The present invention uses graph neural network (GNN) to model the spatial positions of tunnel sensors and their mutual relationships, realizing global structural relationship modeling, which can capture the safety risks of the overall tunnel structure and enhance the comprehensiveness of the assessment.

[0061] 5. The present invention combines the multi-input deep neural network DNN and the gated recurrent unit GRU to jointly analyze the extracted multi-dimensional features, classifies them through the support vector machine SVM, and outputs an accurate tunnel safety assessment level, achieving high-precision safety assessment and providing a reliable basis for tunnel maintenance.

[0062] 6. Through the incremental training strategy, the present invention can continuously update the model according to the safety monitoring data collected in real time, has incremental learning capabilities, and improves the adaptability and long-term stability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0064] Figure 1 This is a flowchart of a tunnel safety assessment method based on deep learning technology provided by an embodiment of the present invention;

[0065] Figure 2 This is a structural diagram of a tunnel safety assessment system based on deep learning technology provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operating processes, but the protection scope of the present invention is not limited to the following embodiments.

[0067] In the present invention, words such as "in one possible embodiment," "exemplary," or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described in the present invention as "in one possible embodiment," "exemplary," or "for example" should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "in one possible embodiment," "exemplary," or "for example" is intended to present the relevant concepts in a concrete manner.

[0068] Figure 1 The following is a flow chart of a tunnel safety assessment method based on deep learning technology provided by an embodiment of the present invention. Figure 1 As shown, a tunnel safety assessment method based on deep learning technology includes:

[0069] S101, collecting safety monitoring data in the tunnel and preprocessing the collected safety monitoring data; the safety monitoring data includes time series data and spatial data, and the preprocessing includes: data cleaning, data denoising, data standardization, and / or data transformation and enhancement;

[0070] S102: performing sequence modeling on the preprocessed time series data based on a long short-term memory network (LSTM) and / or a transformer model, and outputting time series features for representing time dependency and dynamic change trends; performing local spatial feature extraction on the preprocessed spatial data based on a convolutional neural network (CNN); and performing association modeling on the positions and spatial relationships of the sensors in the tunnel used to collect the safety monitoring data based on a graph neural network (GNN), and outputting global structural features.

[0071] S103, performing joint feature extraction on the temporal features, the local spatial features, and the global structural features based on a multi-input deep neural network DNN and / or a gated recurrent unit GRU model, and outputting the joint features; classifying the joint features based on a support vector machine SVM and / or the DNN, and outputting the safety assessment level of the tunnel.

[0072] In this embodiment, the data processing module can be configured with a variety of sensors, such as:

[0073] 1. From the perspective of physical quantity type, it can include:

[0074] Stress / strain sensors, such as fiber grating sensors (FBGs) with high sensitivity and anti-interference capabilities for measuring strain changes in tunnel structures, strain gauge sensors that assess stress changes in structures by measuring material deformation, and piezoelectric sensors for monitoring tiny stress changes, particularly suitable for dynamic stress monitoring;

[0075] Temperature sensors, such as thermocouples, which are widely used for temperature monitoring in high-temperature environments, and RTDs, which are highly stable and accurate, are used for temperature monitoring in high-temperature environments. They can measure temperature contactlessly and are suitable for use in large areas or hazardous environments.

[0076] Displacement / deformation sensors, such as laser displacement sensors for measuring the relative displacement of an object's surface using a laser beam, suitable for high-precision dynamic displacement monitoring, inductive displacement sensors for measuring the displacement of metal objects, suitable for monitoring the deformation of equipment or structures in tunnels, and LVDT displacement meters with high accuracy and stability for measuring the linear displacement of tunnel structures;

[0077] Acceleration sensors, such as MEMS accelerometers used to measure vibration and acceleration changes of equipment or structures in tunnels, and piezoelectric accelerometers used to monitor the vibration of tunnel structures, especially piezoelectric accelerometers that are advantageous in monitoring low-frequency vibrations;

[0078] Position sensors, such as tilt sensors used to detect the tilt angle of tunnel structures to assess their stability, and inertial measurement units (IMUs) that integrate accelerometers, gyroscopes, and other sensors to accurately measure the motion and attitude changes of objects or structures within tunnels.

[0079] Pressure sensors, such as gas pressure sensors that monitor changes in gas or water pressure in tunnels to prevent safety hazards such as flooding or gas accumulation in tunnels, and liquid level sensors that monitor possible water accumulation in tunnels, especially those commonly found in underground tunnels;

[0080] Gas sensors, such as carbon dioxide (CO2) sensors that monitor carbon dioxide concentration in tunnels to prevent hypoxia or harmful gas accumulation, are particularly used for safety monitoring in mines or underground tunnels, methane gas sensors that prevent methane leaks, and oxygen sensors that monitor oxygen concentration inside tunnels to ensure appropriate ventilation conditions; etc.

[0081] 2. From the perspective of working principle type, it can include:

[0082] Electrical sensors, such as resistive sensors that measure changes in resistance caused by material deformation, capacitive sensors that measure distance, displacement, or humidity by detecting changes in capacitance, and inductive sensors that use changes in inductance to sense the displacement or physical state of an object;

[0083] Optical sensors, such as fiber optic sensors that measure physical quantities such as strain and temperature using fiber Bragg grating or fiber Bragg sensors, and laser sensors that perform contactless measurements using laser beams;

[0084] Piezoelectric sensors, such as piezoelectric accelerometers for monitoring structural vibration and dynamic stress changes, and piezoelectric sensors suitable for high-frequency vibration monitoring and stress assessment;

[0085] Micro-electromechanical systems (MEMS) sensors, such as MEMS accelerometers and gyroscopes, are used for monitoring tunnel vibration and motion, featuring high sensitivity and miniaturization.

[0086] Chemical sensors, such as gas sensors that monitor the concentrations of gases such as oxygen, carbon dioxide, ammonia, and methane and are used for real-time assessment of tunnel ventilation conditions; etc.

[0087] 3. From the perspective of application scenario types, it can include:

[0088] Structural health monitoring sensors are used on the external and internal structures of tunnels to monitor the stability, stress and deformation of the overall structure; environmental monitoring sensors are used to monitor the climate, humidity, temperature, gas composition and other environmental conditions in the tunnel to ensure the safety of tunnel traffic; dynamic monitoring sensors monitor dynamic changes in the tunnel, such as the operating status of equipment, structural vibration, and the flow of vehicles or personnel; hydrological monitoring sensors monitor the water level, flow and pressure of accumulated water in the tunnel to prevent problems such as seepage and flooding; and so on.

[0089] It should be noted that the above description is merely an exemplary description, and in actual applications, settings can be made according to different user needs, and this embodiment does not limit this.

[0090] Based on the above sensors, time series data and spatial data can be collected as safety monitoring data.

[0091] Specifically, time series data such as stress data, displacement data, vibration and dynamic data, hydrological data; spatial data such as temperature data, gas concentration data.

[0092] Among them, stress data is usually used to analyze the mechanical properties of tunnel structures and evaluate the bearing capacity and safety status of tunnels under different working conditions. It can include: axial stress, that is, measuring the stress changes in the axial direction of the tunnel, which is used to evaluate the bearing condition of the tunnel structure under longitudinal load; radial stress, that is, measuring the stress in the radial direction of the tunnel, evaluating the deformation of the tunnel shell and its response to the surrounding rock pressure; shear stress, that is, analyzing the shear force between the tunnel walls or lining layers to help identify potential slip or damage risks; bending stress, that is, analyzing the stress generated when the tunnel structure is subjected to bending load, especially the impact on the tunnel arch structure; internal force distribution, that is, evaluating the stress distribution at different positions in the tunnel, which helps analyze the mechanical behavior and key weaknesses of the tunnel; strain, that is, measuring the local deformation of the tunnel structure through strain gauges or fiber optic sensors, which can be used to infer the stress state and perform structural health monitoring; and so on.

[0093] Displacement data is used to monitor the deformation, displacement, and relative movement of tunnel structures, helping to determine whether irregular displacement or deformation has occurred in the tunnel and assessing its stability. It may include: longitudinal displacement, which monitors the deformation of the tunnel along its longitudinal direction, especially the extension or compression that may occur after the tunnel is subjected to stress; radial displacement, which measures the radial displacement of the tunnel surrounding rock and structural walls, especially the expansion or contraction that may occur in the tunnel walls under external loads; vertical displacement, which measures the deformation of the top or bottom of the tunnel under pressure, especially when the tunnel settles or the pressure of the overlying rock changes; horizontal displacement, which monitors the horizontal displacement of the left and right sides of the tunnel or around the structure, especially the displacement caused by earthquakes or other external forces; displacement rate, which measures the rate of change of displacement in the tunnel structure or surrounding environment to identify potential deformation trends; local displacement, which monitors displacement in specific areas (such as cracks, joints, weak areas, etc.) to assess the risk of local structural deformation; real-time displacement change, which monitors the displacement change of the tunnel in real time using equipment such as laser displacement meters, LVDTs (linear variable differential transformers), and fiber optic sensors.

[0094] Vibration and dynamic data are used to monitor the dynamic behavior of the environment inside and outside the tunnel, especially its response to tunnel traffic, external earthquakes or blasting. They can include: acceleration data, which monitors the acceleration changes of tunnel structures or equipment, especially the vibration response when the tunnel is subjected to traffic, earthquakes, etc.; vibration frequency, which analyzes the relationship between the tunnel's natural frequency and external frequency to help identify the frequency range that may cause resonance; vibration amplitude, which monitors the vibration amplitude occurring in the tunnel to assess whether the structure is affected by external vibrations; seismic response data, which monitors the vibration response and deformation of the tunnel structure, especially for external impact forces such as earthquakes; and so on.

[0095] Hydrological data is used to monitor water levels, water accumulation, and humidity in tunnels to ensure there is no water accumulation or leakage in the tunnel. It can include: water level data, which monitors water accumulation inside the tunnel or in the surrounding environment, especially during rainy seasons or when groundwater levels fluctuate, to assess whether there are water leakage problems in the tunnel; water pressure data, which measures water pressure inside the tunnel or in the surrounding rock, which is particularly important in preventing floods and evaluating the effectiveness of waterproofing measures; humidity data, which monitors humidity inside or outside the tunnel, especially for real-time monitoring of environmental humidity that may affect the tunnel structure;

[0096] Temperature data reflects the thermodynamic state of the environment inside and outside the tunnel, affecting the expansion and contraction of tunnel materials and the impact of the surrounding environment on the structure. It can include: temperature inside the tunnel, that is, monitoring the temperature distribution inside the tunnel, usually monitored by temperature sensors (such as thermocouples, RTDs); temperature outside the tunnel, that is, monitoring temperature changes outside the tunnel, especially when the seasons change and the external climate conditions change; material temperature, that is, monitoring the temperature of the tunnel lining or other structural materials to help analyze material expansion and contraction and their impact on the tunnel structure; temperature gradient, that is, calculating the temperature difference inside and outside the tunnel, especially in deep tunnels or extreme weather conditions, the temperature gradient has a significant impact on structural stability; ambient temperature change, that is, monitoring the changes in the surrounding temperature to evaluate the potential impact on the temperature distribution and structure inside the tunnel; and so on.

[0097] Gas concentration data in tunnels is crucial for ventilation systems, personnel safety, and environmental monitoring. Gas concentration data can include: oxygen concentration, which monitors the oxygen content in the tunnel to ensure that the air quality in the tunnel meets safety standards and avoids safety hazards such as suffocation; carbon dioxide concentration, which monitors the concentration of carbon dioxide in the tunnel to prevent accidents caused by hypoxia or gas accumulation; methane concentration, which monitors the concentration of methane gas to prevent explosion risks, especially in coal mine tunnels or underground tunnels; and so on.

[0098] After the safety monitoring data in the tunnel is collected, the collected safety monitoring data may be pre-processed.

[0099] Exemplarily, the preprocessing of the safety monitoring data may include data cleaning, data denoising, data standardization, and / or data transformation and enhancement.

[0100] Specifically, the goal of data cleaning is to remove invalid or erroneous data, fill missing values, and ensure that the data quality meets the requirements of subsequent analysis. It can include:

[0101] 1. Missing value processing, for example, deletion method: directly delete samples or features containing missing values, suitable for situations with fewer missing values; mean / median filling: for numerical data, the mean or median can be used to fill missing values ​​to reduce deviation; interpolation method: use linear interpolation, spline interpolation and other methods to fill missing values ​​in continuous data; forward / backward filling: for time series data, the previous or next data point can be used to fill missing values, which is often used for sensor data; regression filling: use other variables for regression analysis to predict and fill missing data; and so on.

[0102] 2. Outlier processing, for example, box plot method: Identify and remove outliers through box plots. Box plots can identify abnormal points in data distribution, which are usually regarded as "outliers"; standard deviation unit distance Z-score method: By calculating the Z-score of each data point, points that exceed the set threshold are regarded as outliers and eliminated; interquartile range (IQR) method: By calculating the IQR of the data, data points below the lower limit or above the upper limit are regarded as outliers; model-based anomaly detection: methods such as isolation forest or support vector machine (SVM) identify abnormal patterns in data; and so on.

[0103] 3. Data consistency check, for example, logical consistency check: ensuring that the values ​​in the data are physically reasonable, for example, the temperature in the tunnel should not be negative, and the reading of the displacement sensor should not exceed the maximum displacement range of the structure; time consistency check: ensuring that the timestamps of time series data are continuous and non-duplicate to avoid data misalignment or duplication; etc.

[0104] The goal of data denoising is to eliminate or reduce the noise introduced during the data collection process, thereby making the data more accurate and reliable. It can include:

[0105] 1. Filtering methods, such as low-pass filter: removes high-frequency noise and retains the low-frequency components of the signal; suitable for vibration and dynamic monitoring data; high-pass filter: removes low-frequency noise and retains the high-frequency components of the signal; suitable for removing slowly changing interference; band-pass filter: retains only signals within a specific frequency range and removes high- and low-frequency noise; Kalman filter: suitable for noise suppression in dynamic systems, and can predict and correct noise through system models; etc.

[0106] 2. Use wavelet transform to decompose the signal and perform denoising at different frequency levels, which can retain important features while removing noise; it is often used to denoise complex signals such as strain signals and vibration signals.

[0107] 3. Use principal component analysis (PCA) to reduce the dimensionality of high-dimensional data and remove the noise signal by removing the principal components with smaller variance, thereby retaining the main features in the data.

[0108] 4. Use the autoencoder model in deep learning to remove noise from the data and restore the true signal by training noise-free input and output pairs.

[0109] 5. Perform sliding average processing on the data to smooth short-term fluctuations and remove rapidly changing noise; suitable for smoothing time series data.

[0110] The purpose of data normalization is to convert the data into a unified scale to facilitate subsequent analysis and model training, and to avoid the influence of different data dimensions on the results. It can include: Z-score normalization, minimum-maximum normalization, maximum absolute value normalization, normalization based on median and IQR, logarithmic transformation unit vector normalization, etc.

[0111] In some cases, in addition to standardization, data may also need to be transformed or enhanced to suit specific model requirements. This can include: performing sliding window processing on time series data to create fixed-length time series segments for model learning and analysis; rotating, translating, adding noise, and other operations on sensor data to generate more samples and enhance the robustness of the model; performing Fourier transform or wavelet transform on the signal to convert it to the frequency domain for analysis, especially for vibration signals and dynamic monitoring data; and so on.

[0112] Then, the preprocessed time series data is subjected to sequence modeling based on a long short-term memory network (LSTM) and / or a transformer model, and time series features for representing time dependency and dynamic change trends are output;

[0113] Performing local spatial feature extraction on the preprocessed spatial data based on a convolutional neural network (CNN);

[0114] Based on the graph neural network (GNN), the positions and spatial relationships of the sensors used to collect the safety monitoring data in the tunnel are associated and modeled to output global structural features.

[0115] In a possible embodiment, in an optional embodiment, performing sequence modeling on the preprocessed time series data based on LSTM includes:

[0116] The input of the LSTM is the time series data X t ={x1,x2,…,x t}, and output the hidden state H based on the following formula LSTM ={h1,h2,…,h t}:

[0117] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0118] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0119]

[0120] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0121] h t =o t ☉tanh(C t )

[0122] Among them, x t is the input of time series data at time t, f t is the output of the forget gate, i t is the output of the input gate, is the candidate cell state, C t To update the cell state, o t is the output of the output gate, h t is the hidden state at time t, σ is the sigmoid activation function, ⊙ represents element-wise multiplication, W f ,W i ,W C ,W o is the weight matrix, b f ,b i ,b C ,b o is the bias term.

[0123] Recurrent neural networks (RNNs), LSTMs, and GRUs are suitable for processing data with temporal dependencies, such as time-varying data such as stress, temperature, and displacement in tunnels. By learning long-term dependencies in time series data, the model can predict future dynamic trends. Furthermore, based on historical time series data, RNNs, LSTMs, and GRUs can be used to detect abnormal changes in tunnels at specific points in time or over a period of time, thereby predicting potential risks. Therefore, RNNs, LSTMs, and GRUs can effectively capture both long-term and short-term dependencies in time series, making them suitable for processing dynamic changes. They have excellent modeling capabilities for temporal correlations in tunnel data (such as trends in stress and displacement), making them suitable for processing stress data, displacement data, vibration and dynamic data, hydrological data, gas concentration data, and especially time series data.

[0124] In an optional embodiment, performing sequence modeling on the preprocessed time series data based on a Transformer model includes:

[0125] The input of the Transformer model is the time series data X t ={x1,x2,…,xt}, and output the representation Y after self-attention mechanism processing based on the following self-attention mechanism formula t ={y1,y2,…,y t}:

[0126] Q t =W Q ·X t , K t =W K ·X t , V t =W V ·X t

[0127]

[0128] Y t =Attention(Q t ,K t , V t )

[0129] Among them, Q t is the query matrix, K t is the bond matrix, V t is the value matrix, W Q is the weight matrix corresponding to the query matrix, W K is the weight matrix corresponding to the key matrix, W V is the weight matrix corresponding to the value matrix, d k is the dimension of the key vector, and softmax(·) is the normalization operation.

[0130] The Transformer model, through its self-attention mechanism, can capture long-range dependencies, making it ideal for processing time-series data with long-term dependencies. In tunnel safety assessments, different sensor data may exhibit complex temporal dependencies, and the Transformer model can effectively model these dependencies. Furthermore, the Transformer model can handle multiple different types of input (such as sensor data from different sources) and combine this information for modeling, making it suitable for comprehensive analysis of multi-dimensional data (such as temperature, vibration, and stress). Therefore, the Transformer model's efficient sequence modeling capabilities are particularly well-suited for processing long-sequence data. Its self-attention mechanism can focus on important features, making it suitable for multimodal data fusion and for processing stress data, displacement data, vibration and dynamic data, hydrological data, and gas concentration data.

[0131] In an optional embodiment, performing local spatial feature extraction on the pre-processed spatial data based on a convolutional neural network (CNN) includes:

[0132] The input of the convolution operation of the CNN is the spatial data X i ={x1,x2,…,x i}, and output the local spatial feature Y based on the calculation formula of the following convolution operation i ={y1,y2,…,y i}:

[0133]

[0134] Among them, j represents the neighbor node of node i, k is the total number of neighbor nodes of each node i, X i+j is the local area of ​​the input spatial data, W j is the convolution kernel, b is the bias term, and f(·) is the activation function.

[0135] CNN is very good at extracting local features, especially when processing spatial data (such as images or two-dimensional grids). If the monitoring data of the tunnel can be represented in the form of images, heat maps or 2D / 3D grids, CNN can be used to automatically extract local and global features of the tunnel structure, which is suitable for processing spatial data. In time series data processing, CNN can be applied to the convolution operation of time series data to identify the change pattern at different time steps and capture trends. Therefore, CNN is suitable for processing temperature data, displacement data, vibration and dynamic data, hydrological data, gas concentration data and other data, especially image and grid data. In addition, CNN can be used for multi-channel input and can process the combination of multiple sensor data.

[0136] In an optional embodiment, the graph neural network (GNN)-based modeling performs correlation modeling on the positions and spatial relationships of the sensors used to collect the safety monitoring data in the tunnel, and outputs global structural features, including:

[0137] The input of the GNN is the location node feature H = {h1,h2,…,h n}, and output the global structural feature H'={h'1,h'2,…,h' based on the calculation formula of the following graph convolution operation n}:

[0138]

[0139] Among them, h i is the feature of the i-th node, N(i) is the set of neighbor nodes of the i-th node, c ij is the normalization factor, σ is the sigmoid activation function, W is the weight matrix, and b is the bias term.

[0140] A tunnel can be viewed as a graph structure, with sensor locations as nodes and edges between nodes representing the relationships or spatial structure between sensors. GNNs can learn the features of nodes and edges in the graph to infer the overall safety status of the tunnel. Through graph convolution, GNNs can capture changes in local and global regions, helping to assess the safety risks of the entire tunnel. Therefore, GNNs are capable of processing structured spatial data, especially when sensor data distribution is complex. For multi-sensor data, GNNs can consider the spatial relationships between sensors and effectively extract interaction features between nodes. If the tunnel monitoring data involves spatial structure (for example, sensors are distributed at different locations in the tunnel), GNNs are well suited to processing graph-structured data.

[0141] In an optional embodiment, the performing joint feature extraction on the temporal features, the local spatial features, and the global structural features based on a multi-input deep neural network DNN and outputting the joint features includes:

[0142] Based on the following formula, the timing feature f time , the local spatial feature f local , and the global structural feature f global Splice and get the overall input vector x fusion :

[0143] x fusion =concat(f time ,f local ,f global )

[0144] The DNN includes L fully connected layers, and the output of the lth layer is expressed by the following formula:

[0145] h (l) =σ(W (l) h (l-1) +b (l) ),l=1,...,L

[0146] Among them, h (0) =x fusion , W (l) is the weight matrix, b (l) is the bias term, σ is the sigmoid activation function, and the joint feature is f joint =h (L) .

[0147] In an optional embodiment, the gated recurrent unit (GRU) model is used to perform joint feature extraction on the temporal features, the local spatial features, and the global structural features, and output a fused feature vector, including:

[0148] Based on the following formula, the timing feature f time , the local spatial feature f local , and the global structural feature f global Splice and get the overall input vector x fusion :

[0149] x fusion =concat(f time , f local , f global )

[0150] The reset gate r of the GRU t , update gate z t , candidate hidden states The final hidden state h t They are expressed by the following formulas:

[0151] r t =σ(W r ·[h t-1 ,x t ]+b r )

[0152] z t =σ(W z ·[h t-1 ,x t ]+b z )

[0153]

[0154] Among them, W r ,W z ,W h is the weight matrix, b r ,b z ,b h is the bias term, σ is the sigmoid activation function, and the joint feature is f joint =h t .

[0155] In an optional embodiment, the classifying the joint features based on a support vector machine (SVM) and outputting the security assessment level of the tunnel includes:

[0156] The input of the SVM is the joint feature f joint , and based on the following classification decision function f(f joint ) Output the security assessment level of the tunnel:

[0157] f(f joint )=w T f joint +b

[0158] Among them, w is the weight matrix, w T is the transpose of vector w, and b is the bias term.

[0159] It should be noted that the above is only an exemplary description. In actual applications, only one of the above methods can be used for feature processing according to different user needs, or a combination of two or more methods can be used for feature processing. This embodiment does not limit this.

[0160] In an optional embodiment, the method further includes:

[0161] An incremental training set is updated according to the security monitoring data and the security assessment level, and incremental training is performed on the LSTM, the Transformer model, the CNN, the GNN, the DNN, the GRU, and / or the SVM based on the incremental training set and a preset reinforcement learning strategy.

[0162] To improve the adaptability of the assessment system, an incremental learning mechanism can be introduced. Whenever new monitoring data (such as stress, temperature, and displacement changes) is collected, the system will fine-tune the existing model based on the new data. For example, at regular intervals, new monitoring data is uploaded to a cloud-based data processing center via a wireless network or data transmission module for data updates. Based on the new data, the training set is incrementally updated, and reinforcement learning strategies (such as Q-learning and deep Q networks (DQNs)) are used to incrementally train models such as LSTM, Transformer, CNN, and GNN. This allows the model to maintain high accuracy and stability as it continuously receives new data. Consequently, the model is regularly verified and adjusted to ensure the accuracy and reliability of tunnel safety assessments under different conditions. Through this incremental learning strategy, the system can continuously optimize the assessment model based on long-term tunnel monitoring data, adapting to the dynamic characteristics of tunnel structure and external environmental changes, and further improving the long-term stability and applicability of the assessment.

[0163] In addition, those skilled in the art may also adopt other deep learning models for feature processing based on this, such as feedforward neural networks (FFNN), recurrent neural networks (RNN), generative adversarial networks (GAN), autoencoders (AE), reinforcement learning (RL), etc., which are not limited in this embodiment.

[0164] For example, principal component analysis (PC) can be used to reduce the dimension of joint features to reduce computational complexity, and random forests can be used to optimize classification results and improve the model's generalization ability. When outputting the assessment level, Bayesian inference or fuzzy logic methods can be combined to output a confidence value to indicate the reliability of the assessment result. Tunnel monitoring data is combined with historical safety data, and transfer learning methods are used to pre-train the model to improve its adaptability in new tunnel scenarios. Based on the tunnel's geographic information system (GIS) data, the tunnel's spatial distribution information is integrated with sensor data, and spatial convolution operations are used to improve the dynamic analysis capabilities of the tunnel's local structure.

[0165] Based on the above-mentioned method, the present invention integrates multi-dimensional information, such as time series data, spatial data, and the spatial position relationships of sensors, into a comprehensive model. This integrates multi-source data and comprehensively and accurately reflects the safety status of tunnels, avoiding the assessment errors caused by single data types in traditional methods. The present invention uses a long short-term memory (LSTM) network and a transformer model to model time series data, achieving efficient dynamic modeling and prediction capabilities. This effectively captures the dynamic changes in tunnel structures, thereby improving the ability to predict future risks. The present invention utilizes a convolutional neural network (CNN) to extract features from spatial data, achieving precise spatial feature extraction and automatically identifying potential local structural issues within the tunnel, thereby improving the accuracy of spatial information processing. The present invention uses a graph neural network (GNN) to model the spatial positions of tunnel sensors and their interrelationships, achieving global structural relationship modeling, capturing safety risks across the entire tunnel structure and enhancing the comprehensiveness of the assessment. The present invention combines a multi-input deep neural network (DNN) and a gated recurrent unit (GRU) to jointly analyze the extracted multi-dimensional features, classify them using a support vector machine (SVM), and output an accurate tunnel safety assessment level. This achieves highly accurate safety assessments and provides a reliable basis for tunnel maintenance. Through the incremental training strategy, the present invention can continuously update the model according to the safety monitoring data collected in real time, has incremental learning capabilities, and improves the adaptability and long-term stability of the model.

[0166] Figure 2 This is a schematic diagram of the structure of a tunnel safety assessment system based on deep learning technology provided by an embodiment of the present invention. Figure 2 The tunnel safety assessment system 10 based on deep learning technology is characterized by comprising: a data processing module 101, a feature processing module 102, and a level assessment module 103;

[0167] The data processing module 101 is used to collect safety monitoring data in the tunnel and pre-process the collected safety monitoring data; the safety monitoring data includes time series data and spatial data, and the pre-processing includes: data cleaning, data denoising, data standardization, and / or data transformation and enhancement;

[0168] The feature processing module 102 is used to perform sequence modeling on the pre-processed time series data based on a long short-term memory network (LSTM) and / or a transformer model, and output time series features for representing time dependency and dynamic change trends; extract local spatial features from the pre-processed spatial data based on a convolutional neural network (CNN); and perform correlation modeling on the positions and spatial relationships of the sensors in the tunnel used to collect the safety monitoring data based on a graph neural network (GNN), and output global structural features;

[0169] The level assessment module 103 is used to perform joint feature extraction on the temporal features, the local spatial features and the global structural features based on a multi-input deep neural network DNN and / or a gated recurrent unit GRU model, and output the joint features; classify the joint features based on a support vector machine SVM and / or the DNN, and output the security assessment level of the tunnel.

[0170] Furthermore, it should be noted that the present invention may be provided as a method, apparatus, or computer program product. Thus, embodiments of the present invention may take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention may take the form of a computer program product embodied on one or more computer-usable storage media containing computer-usable program code.

[0171] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the process in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0172] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device, so that a series of operation steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for implementing the process in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0173] It should also be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. The terms "comprises," "includes," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or terminal device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal device. In the absence of further restrictions, an element defined by the phrase "comprises a ..." does not exclude the presence of additional identical elements in the process, method, article, or terminal device that includes the element.

[0174] Finally, it should be noted that the above is a preferred embodiment of the present invention. It should be noted that although the preferred embodiment of the present invention has been described, it is clear that those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles of the present invention. Such improvements and modifications should also be considered as within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including the preferred embodiment and all changes and modifications that fall within the scope of the embodiments of the present invention.

Claims

1. A tunnel safety assessment method based on deep learning technology, characterized in that: include: Collecting safety monitoring data in the tunnel and preprocessing the collected safety monitoring data; The safety monitoring data includes time series data and spatial data, and the preprocessing includes: data cleaning, data denoising, data standardization, and / or data transformation and enhancement; Performing sequence modeling on the preprocessed time series data based on a long short-term memory network (LSTM) and / or a transformer model to output time series features for representing time dependency and dynamic change trends; extracting local spatial features from the preprocessed spatial data based on a convolutional neural network (CNN); and performing association modeling on the positions and spatial relationships of the sensors in the tunnel used to collect the safety monitoring data based on a graph neural network (GNN) to output global structural features; Based on a multi-input deep neural network DNN and / or a gated recurrent unit GRU model, joint feature extraction is performed on the temporal features, the local spatial features, and the global structural features, and the joint features are output; based on a support vector machine SVM and / or the DNN, the joint features are classified, and the safety assessment level of the tunnel is output.

2. The tunnel safety assessment method based on deep learning technology according to claim 1, characterized in that: The performing sequence modeling on the pre-processed time series data based on LSTM includes: The input of the LSTM is the time series data X t ={x1,x2,…,x t }, and output the hidden state H based on the following formula LSTM ={h1,h2,…,h t }: f t =σ(W f ·[h t-1 ,x t ]+b f ) =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ☉tanh(C t ) Among them, x t is the input of time series data at time t, f t is the output of the forget gate, i t is the output of the input gate, is the candidate cell state, C t To update the cell state, o t is the output of the output gate, h t is the hidden state at time t, σ is the sigmoid activation function, ⊙ represents element-wise multiplication, W f ,W i ,W C ,W o is the weight matrix, b f ,b i ,b C ,b o is the bias term.

3. The tunnel safety assessment method based on deep learning technology according to claim 1, characterized in that: The performing sequence modeling on the pre-processed time series data based on the Transformer model includes: The input of the Transformer model is the time series data X t ={x1,x2,…,x t }, and output the representation Y after self-attention mechanism processing based on the following self-attention mechanism formula t ={y1,y2,…,y t }: Q t =W Q ·X t ,K t =W K ·X t ,V t =W V ·X t Y t =Attention(Q t ,K t ,V t ) Among them, Q t is the query matrix, K t is the bond matrix, V t is the value matrix, W Q is the weight matrix corresponding to the query matrix, W K is the weight matrix corresponding to the key matrix, W V is the weight matrix corresponding to the value matrix, d k is the dimension of the key vector, and softmax(·) is the normalization operation.

4. The tunnel safety assessment method based on deep learning technology according to claim 1, characterized in that: The local spatial features of the pre-processed spatial data are extracted based on a convolutional neural network (CNN), including: The input of the convolution operation of the CNN is the spatial data X i ={x1,x2,…,x i }, and output the local spatial feature Y based on the calculation formula of the following convolution operation i ={y1,y2,…,y i }: Among them, j represents the neighbor node of node i, k is the total number of neighbor nodes of each node i, X i+j is the local area of ​​the input spatial data, W j is the convolution kernel, b is the bias term, and f(i) is the activation function.

5. The tunnel safety assessment method based on deep learning technology according to claim 1, characterized in that: The graph neural network (GNN) is used to perform correlation modeling on the positions and spatial relationships of the sensors in the tunnel used to collect the safety monitoring data, and output global structural features, including: The input of the GNN is the location node feature H = {h1,h2,…,h n }, and output the global structural feature H'={h'1,h'2,…,h' based on the calculation formula of the following graph convolution operation n }: Among them, h i is the feature of the i-th node, N(i) is the set of neighbor nodes of the i-th node, c ij is the normalization factor, σ is the sigmoid activation function, W is the weight matrix, and b is the bias term.

6. The tunnel safety assessment method based on deep learning technology according to claim 1, characterized in that: The performing joint feature extraction on the temporal features, the local spatial features, and the global structural features based on a multi-input deep neural network DNN and outputting the joint features includes: Based on the following formula, the timing feature f time , the local spatial feature f local , and the global structural feature f global Splice and get the overall input vector x fusion : x fusion =concat(f time ,f local ,f global ) The DNN includes L fully connected layers, and the output of the lth layer is expressed by the following formula: h (l) =σ(W (l) h (l-1) +b (l) ),l=1,...,L Among them, h (0) =x fusion , W (l) is the weight matrix, b (l) is the bias term, σ is the sigmoid activation function, and the joint feature is f joint =h (L) .

7. The tunnel safety assessment method based on deep learning technology according to claim 1, characterized in that: The gated recurrent unit (GRU) model is used to perform joint feature extraction on the temporal features, the local spatial features, and the global structural features, and output a fused feature vector, including: Based on the following formula, the timing feature f time , the local spatial feature f local , and the global structural feature f global Splice and get the overall input vector x fusion : x fusion =concat(f time ,f local ,f global ) The reset gate r of the GRU t , update gate z t , candidate hidden states The final hidden state h t They are expressed by the following formulas: r t =σ(W r ·[h t-1 ,x t ]+b r ) z t =σ(W z ·[h t-1 ,x t ]+b z ) Among them, W r ,W z ,W h is the weight matrix, b r ,b z ,b h is the bias term, σ is the sigmoid activation function, and the joint feature is f joint =h t .

8. The tunnel safety assessment method based on deep learning technology according to claim 1, characterized in that: The classifying the joint features based on a support vector machine (SVM) and outputting a security assessment level of the tunnel includes: The input of the SVM is the joint feature f joint , and based on the following classification decision function f(f joint ) Output the security assessment level of the tunnel: f(f joint )=w T f joint +b Among them, w is the weight matrix, w T is the transpose of vector w, and b is the bias term.

9. The tunnel safety assessment method based on deep learning technology according to claim 1, characterized in that: The method further comprises: An incremental training set is updated according to the security monitoring data and the security assessment level, and incremental training is performed on the LSTM, the Transformer model, the CNN, the GNN, the DNN, the GRU, and / or the SVM based on the incremental training set and a preset reinforcement learning strategy.

10. A tunnel safety assessment system based on deep learning technology, characterized in that: include: Data processing module, feature processing module, and grade assessment module; The data processing module is used to collect safety monitoring data in the tunnel and pre-process the collected safety monitoring data; The safety monitoring data includes time series data and spatial data, and the preprocessing includes: data cleaning, data denoising, data standardization, and / or data transformation and enhancement; The feature processing module is used to perform sequence modeling on the pre-processed time series data based on a long short-term memory network (LSTM) and / or a transformer model, and output time series features for representing time dependency and dynamic change trends; extract local spatial features from the pre-processed spatial data based on a convolutional neural network (CNN); and perform association modeling on the positions and spatial relationships of the sensors in the tunnel used to collect the safety monitoring data based on a graph neural network (GNN), and output global structural features; The grade assessment module is used to perform joint feature extraction on the temporal features, the local spatial features and the global structural features based on a multi-input deep neural network DNN and / or a gated recurrent unit GRU model, and output the joint features; classify the joint features based on a support vector machine SVM and / or the DNN, and output the security assessment grade of the tunnel.

Images