Infrared thermography identification method and system for water gushing sign of fault zone tunnel

Abstract

This invention discloses an infrared thermal imaging identification method and system for water inrush signs in tunnels along fault zones, relating to the field of tunnel engineering monitoring and disaster early warning technology. The method includes a multi-source data time-series acquisition step, an infrared temperature field time-series analysis step, a multi-modal identification step for seepage channels, a water inrush precursor feature extraction step, and a closed-loop feedback optimization step. This invention captures the evolution law of low-temperature anomaly areas before water inrush through infrared thermal imaging time-series analysis, and uses a cross-modal attention fusion network to achieve adaptive fusion of infrared and visible light features. It calculates water inrush risk assessment indicators by comprehensively considering multi-dimensional parameters, thereby achieving early warning of water inrush disasters with an advance warning of two to six hours.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering monitoring and disaster early warning technology, specifically to an infrared thermal imaging identification method and system for signs of water inrush in tunnels along fault zones. Background Technology

[0002] With the rapid development of transportation infrastructure construction in my country, tunnel projects traversing complex geological structures are increasing. Fault zones, as weak areas in geological structures, are characterized by fractured rock masses and well-developed joints and fissures, making them high-risk areas for water inrush and mudslide disasters during tunnel construction. Water inrush and mudslide disasters are characterized by their suddenness, destructive power, and difficulty in early warning. Once they occur, they often cause significant casualties and property losses, seriously affecting tunnel construction safety and project progress.

[0003] Chinese invention CN111899288A discloses a method for detecting and identifying water seepage areas in tunnels based on the fusion of infrared and visible light images. This method utilizes an industrial camera to acquire infrared and visible light images of the tunnel interior. The acquired images are preprocessed and registered, and then fused using a generative adversarial network model from deep learning. The fused image is used for target detection of water seepage and removal of abnormal epoxy resin conditions. Finally, an inertial navigation system is used to locate and mark the water seepage area. This method combines the advantages of infrared and visible light images, enabling water seepage detection even in tunnel environments with poor lighting conditions.

[0004] However, the technical solutions of the above inventions have the following shortcomings. They primarily target the detection and identification of areas where leakage has already occurred, falling under the category of post-event detection and failing to provide early warning of water inrush disasters. They employ a single-frame image analysis method, neglecting the temporal evolution of the temperature field and thus failing to capture the gradual changes preceding water inrush. The image fusion method uses generative adversarial networks, which, while improving the quality of the fused images, are not optimized for the task of identifying seepage channels and lack a mechanism for the complementary utilization of infrared thermal features and visible light texture features. Furthermore, a closed-loop feedback mechanism between the early warning model and actual water inrush events is not established, making it impossible to continuously optimize early warning performance based on historical data.

[0005] Identifying signs of water inrush in tunnels along fault zones presents unique challenges. Groundwater in fault zones often seeps into tunnels through complex fracture networks, exhibiting a gradual evolution from point seepage to linear flow and then to surface inrush. Before water inrush occurs, the surface temperature of the seepage area gradually decreases due to continuous groundwater infiltration, forming a low-temperature anomaly zone, which expands as the seepage volume increases. Existing manual observation methods rely on the experience of monitoring personnel, which suffers from strong subjectivity, fatigue, and difficulty in continuous monitoring, easily missing early signs of water inrush. Therefore, there is an urgent need for an intelligent method that can automatically identify precursory features of water inrush and provide early warnings. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an infrared thermal imaging identification method and system for identifying signs of water inrush in tunnels along fault zones. By combining infrared thermal imaging time-series analysis with multimodal feature fusion, the method automatically identifies precursory features of water inrush, enabling early warning of water inrush disasters.

[0007] This invention provides an infrared thermal imaging method for identifying signs of water inrush in tunnels along fault zones, comprising the following steps.

[0008] The multi-source data time-series acquisition steps involve acquiring infrared thermal imaging data of the tunnel face and tunnel walls using an infrared thermal imager, acquiring visible light image data using a visible light camera, acquiring ambient temperature data using a temperature sensor, acquiring groundwater level monitoring data using a water level sensor, spatially registering the infrared thermal imaging data and the visible light image data, performing temperature compensation correction on the infrared thermal imaging data based on the ambient temperature data, and generating a time-synchronized multimodal data sequence.

[0009] The infrared temperature field time series analysis steps involve reconstructing the temperature field from the registered and calibrated infrared thermal imaging data, calculating the temperature gradient field, identifying low-temperature anomaly regions based on a preset temperature anomaly threshold, and extracting the time series evolution characteristics of the low-temperature anomaly regions, including the expansion rate of the temperature anomaly region and the rate of change of the temperature gradient.

[0010] The multimodal identification step for seepage channels involves inputting low-temperature anomaly region and visible light image data into a cross-modal attention fusion network. Temperature distribution features are extracted through the infrared feature branch, and texture features of the humid area are extracted through the visible light feature branch. The temperature distribution features and texture features of the humid area are fused through the cross-modal attention mechanism to output the seepage channel identification results, which include the location of the seepage channel and the seepage intensity level.

[0011] The step of extracting precursor features of water inrush involves calculating water inrush risk assessment indicators based on temporal evolution characteristics and seepage channel identification results, combined with groundwater level monitoring data. The water inrush risk level is determined according to the water inrush risk assessment indicators. When the water inrush risk level exceeds the preset risk threshold, a water inrush early warning signal is generated, and the early warning advance is predicted based on the expansion rate of the temperature anomaly zone.

[0012] The closed-loop feedback optimization step records the actual time and intensity of water inrush, compares the actual time of water inrush with the time of water inrush warning signal issuance, calculates the evaluation indicators of warning accuracy and timeliness, adjusts the preset temperature deviation value and preset risk threshold based on the evaluation indicators, and feeds the optimized threshold parameters back to the infrared temperature field time series analysis step and the water inrush precursor feature extraction step.

[0013] The present invention also provides an infrared thermal imaging identification system for signs of water inrush in tunnels in fracture zones, including a multi-source data time-series acquisition module, an infrared temperature field time-series analysis module, a seepage channel multi-modal identification module, a water inrush precursor feature extraction module, and a closed-loop feedback optimization module. Each module has the same function as the corresponding step in the above method.

[0014] The beneficial effects of this invention are as follows.

[0015] This invention, through time-series analysis of infrared thermal imaging data, can capture the gradual expansion pattern of low-temperature anomaly areas before water inrush occurs, enabling early identification of water inrush signs. The warning lead time can reach two to six hours, providing ample time for construction personnel evacuation and emergency response. This invention employs a cross-modal attention fusion network, fully utilizing temperature distribution information from infrared images and texture details from visible light images. Through an attention mechanism, it achieves adaptive fusion of features from both modalities, improving the accuracy of seepage channel identification by more than 15% compared to single-modal methods. This invention establishes a closed-loop feedback mechanism between the warning results and actual water inrush events, continuously optimizing warning threshold parameters based on historical data. With extended operation time, the accuracy and timeliness of warnings continuously improve. This invention comprehensively considers multi-dimensional parameters such as the expansion rate of the temperature anomaly zone, the rate of change of the temperature gradient, the seepage intensity level, and changes in groundwater level. Through weighted fusion calculation of water inrush risk assessment indicators, this invention improves the comprehensiveness and reliability of risk assessment. Attached Figure Description

[0016] Figure 1 This is a flowchart of the infrared thermal imaging identification method for signs of water inrush in tunnels along fracture zones, as described in this invention.

[0017] Figure 2 This is an architecture diagram of the infrared thermal imaging recognition system for signs of water inrush in tunnels along fault zones, as described in this invention. Detailed Implementation

[0018] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] See Figure 1 This invention provides an infrared thermal imaging method for identifying signs of water inrush in tunnels along fault zones. The method includes a multi-source data time-series acquisition step, an infrared temperature field time-series analysis step, a multi-modal identification step for seepage channels, a water inrush precursor feature extraction step, and a closed-loop feedback optimization step. These five steps form a deeply coupled closed-loop collaborative architecture. The output of each step serves as the key input for the next step, and the output of the closed-loop feedback optimization step inversely influences the threshold parameters of the infrared temperature field time-series analysis step and the water inrush precursor feature extraction step. Through continuous iterative optimization, the early warning performance is continuously improved.

[0020] Step S1: Multi-source data time-series acquisition steps.

[0021] The multi-source data time-series acquisition step is the data foundation of the entire identification method, responsible for acquiring and preprocessing heterogeneous data from multiple sensors. In one embodiment of the present invention, this step mainly consists of five sub-processes: infrared thermal imaging data acquisition, visible light image data acquisition, environmental parameter data acquisition, multimodal data registration, and temperature compensation correction.

[0022] The infrared thermal imaging data acquisition process employs an uncooled infrared thermal imager to continuously scan the tunnel face and tunnel walls. Preferably, the infrared thermal imager has a temperature resolution of no less than 0.05 degrees Celsius, a spatial resolution of no less than 640×480 pixels, and a sampling frame rate set to 1 to 5 frames per second, with the specific frame rate determined based on the tunnel construction progress and geological conditions. The infrared thermal imager is mounted on a movable support, capable of covering the entire tunnel face and the tunnel walls within a range of 20 to 50 meters behind the face. During acquisition, the infrared thermal imager automatically records the acquisition timestamp of each frame for subsequent time-series analysis.

[0023] The visible light image data acquisition process employs an industrial-grade visible light camera to simultaneously acquire texture images of the tunnel surface. Preferably, the visible light camera has a resolution of no less than 1920×1080 pixels, is equipped with a wide-angle lens to cover the same field of view as the infrared thermal imager, and is equipped with LED supplementary lighting to overcome the problem of insufficient illumination inside the tunnel. The visible light camera and the infrared thermal imager use a hardware synchronous triggering method to ensure that the acquisition time of the two images is consistent.

[0024] The environmental parameter data acquisition process includes two parts: ambient temperature acquisition and groundwater level monitoring. Ambient temperature acquisition utilizes high-precision temperature sensors with a measurement accuracy of no less than 0.1 degrees Celsius, deployed at multiple locations near the tunnel face and on the tunnel wall, with a sampling cycle consistent with that of the infrared thermal imager. Groundwater level monitoring employs pressure-type water level sensors, installed in a pre-drilling borehole ahead of the tunnel face, to monitor real-time changes in groundwater levels in the fault zone area, with a sampling cycle set to 1 minute.

[0025] The multimodal data registration process spatially aligns infrared thermal imaging data and visible light image data. Due to differences in the installation positions and lens parameters of the infrared thermal imager and the visible light camera, translation, rotation, and scale transformations exist between the two images. In this embodiment, a registration method based on feature point matching is employed. First, scale-invariant feature points are detected in both the visible light and infrared images. Then, a nearest neighbor matching algorithm is used to establish the correspondence between feature points. A random sampling consensus algorithm is used to eliminate mismatched point pairs. Finally, the affine transformation matrix is ​​calculated using the correctly matched point pairs to perform a geometric transformation on the infrared image to align it with the visible light image. The registered image pairs have pixel-level correspondence, providing a foundation for subsequent multimodal fusion.

[0026] The temperature compensation correction process eliminates the influence of ambient temperature changes on infrared thermal imaging data. The ambient temperature inside the tunnel fluctuates with changes in construction activities, ventilation conditions, and external climate conditions. This fluctuation causes systematic deviations in the surface temperature values ​​measured by the infrared thermal imager. This invention proposes an adaptive compensation method based on ambient temperature, with the following correction formula:

[0027] ,

[0028] in, The corrected temperature value is in degrees Celsius. The original temperature value measured by the infrared thermal imager is in degrees Celsius. These are pixel coordinates; The time of data collection; for The ambient temperature at any given time is measured by a temperature sensor and is expressed in degrees Celsius. For reference to ambient temperature, the temperature is set to the tunnel's design operating temperature or historical average temperature, in degrees Celsius. The environmental temperature influence coefficient ranges from 0.8 to 1.2 and is determined through calibration experiments. The temperature change rate influence coefficient, with a value ranging from 0.1 to 0.5 seconds, is used to compensate for the lag effect of rapid changes in ambient temperature on the measurement. The rate of change of ambient temperature, expressed in degrees Celsius per second, is calculated by the difference in ambient temperature between adjacent moments.

[0029] In a preferred embodiment of the present invention, the environmental temperature influence coefficient The calibration was performed using the following method: During periods of stable tunnel ambient temperature, a dry rock mass area was selected as the calibration target. The difference between the temperature measured by the infrared thermal imager and the ambient temperature was recorded. The average value was calculated through multiple measurements as the baseline deviation. Then, during periods of ambient temperature variation, the ratio of the change in measured temperature to the change in ambient temperature was recorded, and the average value of multiple measurements was taken as the baseline deviation. The calibration value. Temperature change rate influence coefficient. The calibration employs a dynamic response test method, which involves manually controlling rapid changes in ambient temperature and observing the response delay of the infrared thermal imager in measuring temperature. The calibration is determined based on the delay time and the magnitude of the temperature change. The value of .

[0030] The accuracy of multimodal data registration directly affects the effect of subsequent multimodal fusion. In this embodiment, the normalized cross-correlation coefficient (RCC) is used as the indicator to evaluate the registration accuracy, requiring that the RCC of the registered infrared image and the visible light image be no less than 0.85. When the registration accuracy does not meet the requirements, the system automatically performs iterative optimization, increasing the number of feature points or adjusting the matching parameters until the accuracy requirement is met. The registration process also needs to address the resolution difference between the two images. When the resolution of the infrared image is lower than that of the visible light image, a bicubic interpolation method is used to upsample the infrared image, ensuring that the two images have the same spatial resolution.

[0031] In addition, the multi-source data time-series acquisition process also includes a data quality inspection process. This process performs integrity checks and outlier detection on the acquired raw data. Integrity checks verify whether there are missing frames or data loss in the data from each sensor. When missing data is detected, time interpolation is used to complete the missing frames if the number is less than a preset threshold; otherwise, the data for that time period is marked as invalid. Outlier detection uses a statistical method, marking data points that deviate from the mean by more than three standard deviations as outliers, and then repairing them using linear interpolation. The data quality inspection process ensures that the data input into subsequent analysis steps has sufficient reliability, providing a high-quality data foundation for identifying signs of water inrush.

[0032] After registration and temperature compensation correction, the multi-source data time-series acquisition step outputs a time-synchronized multimodal data sequence, including a registered infrared image sequence, a registered visible light image sequence, a corrected ambient temperature sequence, and a pre-processed groundwater level sequence, which serve as inputs for subsequent analysis steps.

[0033] Step S2: Infrared temperature field time series analysis steps.

[0034] The infrared temperature field time-series analysis step receives the registered infrared image sequence and corrected ambient temperature sequence output from the multi-source data time-series acquisition step, and performs temperature field reconstruction, temperature gradient field calculation, low-temperature anomaly region identification, and temporal evolution feature extraction. This step is the core foundation for extracting precursor features of water inrush, providing key features for subsequent water inrush risk assessment by capturing the spatiotemporal evolution patterns of the temperature field.

[0035] The temperature field reconstruction process converts the grayscale or radiance values ​​of the infrared image into physical temperature values, forming a two-dimensional temperature field distribution. In this embodiment, the infrared thermal imager has completed internal calibration, and the output image data can be directly read as temperature values. For cases requiring radiometric correction, Planck's radiation law is used for temperature inversion. After temperature field reconstruction, each frame of the infrared image corresponds to one... A temperature matrix of dimension, where and These represent the height and width of the image, respectively.

[0036] The temperature gradient field calculation process involves spatial differentiation of the reconstructed temperature field to obtain the amplitude and direction information of the temperature gradient. The temperature gradient reflects the rate of temperature change in space; a larger temperature gradient typically exists at the boundary of the seepage channel because the temperature in the seepage area is lower than that of the surrounding dry rock mass. In this embodiment, the Sobel operator is used to calculate the temperature gradient. Let... For position The temperature value at that location indicates the horizontal gradient component. and vertical gradient components The gradient magnitudes were calculated using Sobel convolution kernels. and gradient direction angle The calculation formula is:

[0037] ,

[0038] ,

[0039] in, For position The gradient magnitude at a given point, expressed in degrees Celsius per pixel; For position The gradient direction angle at point , in radians, has a range of values ​​of . to The gradient magnitude is greater than the preset gradient threshold. Pixels with high gradients are marked as high gradient points, corresponding to the boundary locations of seepage channels. Preferably, a preset gradient threshold is used. The initial value is set to 0.5 degrees Celsius per pixel to 2.0 degrees Celsius per pixel, and is adjusted according to the thermal properties of the surrounding rock and the seepage characteristics.

[0040] The low-temperature anomaly region identification process is based on an adaptive threshold segmentation method to identify low-temperature anomaly regions in the temperature field. Before water inrush occurs in a fault zone tunnel, groundwater seeps into the tunnel wall through a network of fractures. The surface temperature of the seepage area is lower than that of the surrounding dry area due to water evaporation and heat conduction, forming a low-temperature anomaly region. This invention proposes a dynamic threshold segmentation method, the calculation process of which is as follows: First, the average temperature of the tunnel wall region in the current frame temperature field is calculated. and temperature standard deviation Then, the dynamic temperature threshold is calculated based on statistical characteristics. :

[0041] ,

[0042] in, This represents the average temperature of the tunnel wall area, in degrees Celsius. This represents the standard deviation of temperature in the tunnel wall region, in degrees Celsius. The standard deviation coefficient, ranging from 1.0 to 3.0, is used to control the stringency of the threshold. The preset temperature deviation value is dynamically adjusted by the closed-loop feedback optimization step, with the initial value set to 0.5 degrees Celsius to 2.0 degrees Celsius.

[0043] Temperature value below dynamic temperature threshold Pixels with abnormal low-temperature activity are marked as such. An 8-connected component analysis is performed on these pixels, and those with areas exceeding a preset minimum area threshold are identified. The connected components are identified as low-temperature anomaly regions. Preferably, a minimum area threshold is preset. The settings are from 100 square pixels to 500 square pixels to filter out noise and isolated points. The results of the low-temperature anomaly region identification are output in the form of a binary mask, where 1 represents a low-temperature anomaly region and 0 represents a normal region.

[0044] The temporal evolution feature extraction process analyzes the changing patterns of low-temperature anomaly regions over time to extract precursor features of water inrush. Before water inrush occurs, low-temperature anomaly regions typically exhibit an evolutionary trend of gradual expansion in area and gradual increase in temperature gradient. This invention extracts two key temporal evolution features: the expansion rate of the temperature anomaly region and the rate of change of the temperature gradient.

[0045] The rate of expansion of the temperature anomaly region reflects the increasing trend of the area of ​​the low-temperature anomaly region over time. Let... for The area of ​​the low-temperature anomaly region at any given time. Given the area at the previous acquisition time, the instantaneous expansion rate is... for:

[0046] ,

[0047] in, for The instantaneous expansion rate at any given moment, measured in square pixels per second; for The area of ​​the low-temperature anomaly region at any given time, expressed in square pixels; The time interval between adjacent acquisition moments is in seconds. To eliminate the influence of measurement noise, a moving average filter is applied to the instantaneous expansion velocity, with the moving window length set to 10 to 30 frames. The smoothed temperature anomaly expansion velocity... Output as one of the precursor features of water inrush.

[0048] The rate of change of the temperature gradient reflects the trend of temperature gradient change over time at the boundary of the seepage channel. Let... for The average gradient magnitude at the boundary of the low-temperature anomaly region at a given time is the rate of change of the temperature gradient. for:

[0049] ,

[0050] in, for The rate of change of temperature gradient at any given time, expressed in degrees Celsius per pixel per second; for The average gradient magnitude at the boundary of the low-temperature anomaly region is obtained by averaging the gradient values ​​of the boundary pixels. When the rate of change of the temperature gradient remains positive, it indicates that the seepage intensity is increasing and the risk of water inrush is rising.

[0051] The infrared temperature field time series analysis step outputs the temperature field reconstruction results, temperature gradient field sequence, low temperature anomaly region mask, and time series evolution characteristics, which are then passed to the seepage channel multimodal identification step and the water inrush precursor feature extraction step, respectively.

[0052] In a preferred embodiment of the present invention, the temperature field reconstruction process further includes thermal inertia compensation. Because rock masses possess a certain degree of thermal inertia, changes in surface temperature lag behind the seepage process, which can affect the timeliness of identifying low-temperature anomaly areas. The present invention employs a compensation method based on a heat conduction model, using the thermal diffusivity of the rock mass and the observation time interval to perform forward prediction compensation on the measured temperature, enabling the temperature field to reflect changes in seepage conditions more promptly.

[0053] In the process of identifying low-temperature anomaly regions, in addition to area threshold filtering, morphological processing is also required to eliminate the influence of noise and isolated points. First, an opening operation is performed on the binary mask to remove small noise areas; then, a closing operation is performed to fill small holes inside the low-temperature anomaly region, making the identification result more complete and continuous. The morphological processing uses a circle as the structural element, with the radius determined based on the image resolution and the typical width of the seepage channel, preferably set to 5 to 15 pixels.

[0054] In the process of extracting temporal evolution features, in addition to the expansion rate of the temperature anomaly region and the rate of change of the temperature gradient, this invention also extracts the shape evolution features of the low-temperature anomaly region. The shape evolution features include the rate of change of the region's perimeter, the rate of change of circularity, and the rate of change of the principal axis direction. These shape features can reflect the expansion direction and pattern of the seepage channel, helping to determine the location and scale of the water inrush. The rate of change of the region's perimeter is obtained by calculating the difference in the perimeter of the low-temperature anomaly region at adjacent times and dividing it by the time interval; circularity is defined as the ratio of area to the square of the perimeter, reflecting the regularity of the region's shape; the principal axis direction is obtained by calculating the region's covariance matrix and obtaining the direction angle of the principal eigenvector.

[0055] Step S3: Multimodal identification of seepage channels.

[0056] The multimodal identification step for seepage channels receives the low-temperature anomaly region mask and temperature field reconstruction results output from the infrared temperature field temporal analysis step, as well as the registered visible light image sequence output from the multi-source data temporal acquisition step. It then identifies the location and intensity of the seepage channels using a cross-modal attention fusion network. This step fully utilizes the temperature distribution information from the infrared images and the texture details from the visible light images, employing deep learning methods to achieve adaptive fusion of the two modal features, thereby improving the accuracy and robustness of seepage channel identification.

[0057] The overall architecture of the cross-modal attention fusion network comprises four components: an infrared feature encoder, a visible light feature encoder, a cross-modal attention module, and a feature decoder. In this embodiment, the network employs an end-to-end training method, with the input being infrared and visible light image patches corresponding to the low-temperature anomaly region, and the output being the segmentation mask and intensity level prediction value of the seepage channel.

[0058] The infrared feature encoder is responsible for extracting temperature distribution features from infrared images. The encoder employs a U-Net-like coding structure, comprising four downsampling stages. Each downsampling stage consists of two convolutional blocks and a max-pooling layer. The convolutional blocks use a combination of 3×3 convolutional kernels, batch normalization, and the ReLU activation function. To enhance feature representation, residual connections are introduced in each downsampling stage. The number of feature channels in each stage is 64, 128, 256, and 512, respectively. The final output feature map of the encoder has a spatial resolution of 1 / 16th of the input image and 512 channels.

[0059] The visible light feature encoder is responsible for extracting texture features of the wetted area from the visible light image. The encoder uses the same network structure as the infrared feature encoder, but with independent network parameters. This design takes into account the different imaging principles and feature distributions of infrared and visible light images, requiring separate learning of feature extraction capabilities for each modality. The input to the visible light feature encoder is a three-channel RGB image, with the number of input channels in the first convolutional layer adjusted accordingly to 3.

[0060] The cross-modal attention module is one of the core innovations of this invention, responsible for fusing infrared and visible light features. Traditional multimodal fusion methods typically employ simple feature concatenation or element-wise addition, failing to adaptively utilize the complementary information of different modalities. The cross-modal attention mechanism proposed in this invention uses infrared features as queries and visible light features as keys and values, achieving adaptive feature fusion through attention calculation. Its calculation process is as follows:

[0061] Let the feature map output by the infrared feature encoder be... The feature map output by the visible light feature encoder is ,in For the number of channels, and This refers to the height and width of the feature map. First, the feature map is flattened into a sequence format. Flattened , Flattened ,in The sequence length is given.

[0062] Query Matrix Key matrix Sum matrix The following results were obtained through linear transformations:

[0063] ,

[0064] ,

[0065] ,

[0066] in, To query the transformation matrix, The key transformation matrix is... The value transformation matrix, and The transformed feature dimension is preferably set to 64.

[0067] Attention weight matrix and cross-modal fusion features The calculation formula is:

[0068] ,

[0069] ,

[0070] in, This is a scaling factor used to prevent the gradient of the softmax function from vanishing due to an excessively large dot product result. The function normalizes each row so that the sum of the attention weights is 1; This is the attention weight matrix, and its elements are... Indicates the first in the infrared feature sequence The position corresponds to the th position in the visible light feature sequence. Attention weights for each position; This is a cross-modal fusion feature.

[0071] Through the aforementioned cross-modal attention calculation, infrared features can adaptively focus on texture information related to temperature anomaly regions in visible light images. When infrared features at a certain location indicate the presence of a low-temperature anomaly, the attention mechanism assigns higher weight to the texture features of the corresponding visible light humidified area, thereby enhancing the discriminative ability to identify seepage channels.

[0072] The feature decoder is responsible for restoring the fused features into a segmentation mask with the same resolution as the input image. The decoder employs a symmetrical upsampling structure, comprising four upsampling stages. Each upsampling stage consists of a transposed convolutional layer, feature concatenation, and two convolutional blocks. The transposed convolution achieves a 2x upsampling, and the feature concatenation joins the features from the current stage with the skip connection features from the corresponding stage in the encoder along the channel dimension, enhancing the recovery of high-resolution details. The final layer of the decoder uses a 1×1 convolution to reduce the number of feature channels to two, corresponding to the water seepage channels and the background, respectively. The softmax function outputs the probability of each pixel belonging to each category.

[0073] Seepage intensity level prediction is another output of multimodal recognition of seepage channels. In this embodiment, seepage intensity is divided into five levels: no seepage, slight seepage, significant seepage, severe seepage, and gushing water. The prediction of seepage intensity level is achieved by adding a global average pooling layer and a fully connected classification layer to the fused features. Global average pooling compresses the fused feature map into a one-dimensional feature vector, the fully connected layer outputs scores for the five categories, and the probability distribution of each level is obtained through a softmax function. The level with the highest probability is taken as the prediction result.

[0074] The multimodal identification step of seepage channels outputs the seepage channel identification results, including the seepage channel location (represented in the form of a segmentation mask), the seepage intensity level (represented by integers 1 to 5), and the visible light wet area detection results. These outputs are passed to the water surge precursor feature extraction step.

[0075] In a preferred embodiment of the present invention, the training of the cross-modal attention fusion network employs a two-stage strategy. The first stage is a pre-training stage, in which the infrared feature encoder and the visible light feature encoder are independently pre-trained, using a single-modal water seepage region segmentation task for supervised learning, enabling the encoders to learn effective feature representations for their respective modalities. The second stage is a joint fine-tuning stage, in which the encoder parameters are fixed, and the cross-modal attention module and feature decoder are trained, using paired infrared and visible light labeled data for supervised learning. This two-stage training strategy can fully utilize single-modal data resources while avoiding the instability issues in the early stages of cross-modal fusion training.

[0076] The loss function for network training consists of two parts: segmentation loss and intensity classification loss. The segmentation loss is a weighted combination of cross-entropy loss and Dice loss. Cross-entropy loss penalizes pixel-by-pixel classification errors, while Dice loss optimizes the overlap between the segmented and labeled regions. Let the segmentation prediction be... The actual label is Then the combination loss The calculation formula is:

[0077] ,

[0078] in, Total number of pixels; For the first The actual label of each pixel, with a value of 0 or 1; For the first The probability of a pixel being predicted as positive; The weighting coefficients for the Dice loss are preferably set to 0.5 to 1.0. The intensity classification loss uses the standard multi-class cross-entropy loss. The total loss is the weighted sum of the segmentation loss and the intensity classification loss, with a preferred weight ratio of 4:1.

[0079] The network training optimizer uses the Adam optimization algorithm with an initial learning rate of 0.001. A cosine annealing learning rate scheduling strategy is employed, adjusting the learning rate based on a cosine function after each training epoch. The training dataset contains 1000 pairs of paired infrared and visible light images, with annotations including pixel-level segmentation masks of seepage areas and seepage intensity level labels. Data augmentation methods include random cropping, random flipping, random rotation, and random brightness adjustment to improve the model's generalization ability. The training epochs are set to 100, with a batch size of 8, and the training time on an NVIDIA RTX 3080 GPU is approximately 8 hours.

[0080] Step S4: Extraction of features indicating impending water inrush.

[0081] The water inrush precursor feature extraction step receives the time-series evolution features output from the infrared temperature field time-series analysis step, the seepage channel identification results output from the seepage channel multi-modal identification step, and the groundwater level monitoring data output from the multi-source data time-series acquisition step. Through multi-parameter fusion, it calculates water inrush risk assessment indicators, determines the water inrush risk level, and generates a water inrush early warning signal. This step is a crucial link in achieving early warning of water inrushes, accurately determining the likelihood and urgency of water inrush occurrence through comprehensive analysis of multi-dimensional precursor features.

[0082] The calculation of the water inrush risk assessment index employs a multi-parameter weighted fusion method. Water inrush is typically preceded by various precursory phenomena, including expansion of the low-temperature anomaly area, increased temperature gradient, increased seepage intensity, and rising groundwater levels. A single parameter is insufficient to comprehensively reflect the water inrush risk; therefore, this invention normalizes multiple parameters and then performs a weighted summation to obtain a comprehensive water inrush risk assessment index. :

[0083] ,

[0084] in, This is an indicator for assessing the risk of water inrush, with a value ranging from 0 to 1. The higher the value, the higher the risk of water inrush. This represents the normalized expansion rate of the temperature anomaly region. This represents the normalized rate of change of the temperature gradient. This represents the normalized permeability intensity rating; This represents the normalized rate of change in groundwater level. , , , Here are the weighting coefficients for each parameter, satisfying... .

[0085] The normalization of each parameter adopts the minimum-maximum normalization method. Taking the expansion rate of the temperature anomaly zone as an example, the normalization formula is:

[0086] ,

[0087] in, The current rate of expansion of the temperature anomaly zone; and These are the minimum and maximum expansion speeds, determined based on historical data. The normalization methods for other parameters are similar.

[0088] Weighting coefficient to The weights are determined using a correlation analysis method based on historical water inrush events. In the initial stage of system operation, all weight coefficients are set to an equal value of 0.25. As the system accumulates historical data, the weight coefficients are adjusted by calculating the Pearson correlation coefficient between each parameter and the actual occurrence of water inrush events, giving greater weight to parameters with high correlation. Preferably, when the amount of historical data reaches a preset threshold (e.g., 10 water inrush events), an adaptive weight update mechanism is activated.

[0089] The determination of the water inrush risk level is based on water inrush risk assessment indicators. Comparison with preset risk thresholds. This invention classifies water inrush risk into four levels: low risk, medium risk, high risk, and extremely high risk. Let the three preset risk thresholds be... , and ,satisfy The risk level determination rule is as follows: when When it is low risk; when At that time, it was considered a medium-risk period; when At that time, it was high-risk; when This is considered extremely high risk. Preferably, a preset risk threshold is used. , , The initial values ​​are set to 0.3, 0.5 and 0.7 respectively, and are dynamically adjusted by the closed-loop feedback optimization step according to the actual water inflow.

[0090] The water inrush warning signal is triggered when the water inrush risk level exceeds a preset trigger level. In this embodiment, the system automatically generates a water inrush warning signal when the water inrush risk level reaches high risk or extremely high risk. The warning signal includes information such as the current water inrush risk level, risk assessment index value, location of the seepage channel, and warning lead time, and is notified to construction management personnel through audible and visual alarm devices and mobile terminal push notifications.

[0091] The prediction of the early warning lead time is based on the analysis of the expansion trend of the low temperature anomaly region. It is assumed that the area of ​​the low temperature anomaly region will grow linearly at the current expansion rate, reaching a preset critical area. If a sudden water inrush occurs, the advance warning time will be [not specified]. The initial estimate is:

[0092] ,

[0093] in, This is an initial estimate of the lead time for early warning, in hours. The critical area threshold for the low-temperature anomaly region is determined based on historical water inrush event data, and the unit is square pixels. This represents the area of ​​the low-temperature anomaly region at the current moment, expressed in square pixels. This represents the current rate of expansion of the temperature anomaly zone, expressed in square pixels per hour.

[0094] The initial estimates need to be revised based on the trend of groundwater level changes. When the groundwater level is rising, the time of water inrush may be earlier, requiring a reduction in the warning lead time; when the groundwater level is stable or falling, the time of water inrush may be delayed, allowing for an appropriate increase in the warning lead time. The revised warning lead time is as follows. for:

[0095] ,

[0096] in, This represents the current rate of change in groundwater level; a positive value indicates an increase, and a negative value indicates a decrease. The maximum historical rate of change of groundwater level is used for normalization; The correction factor ranges from 0.1 to 0.3. After correction, the lead time for early warning is typically 2 to 6 hours, providing sufficient time for the evacuation of construction personnel and emergency response.

[0097] The step of extracting precursor features of water inrush outputs the water inrush risk level, water inrush warning signal, and warning lead time, which are used for construction safety management decision-making and closed-loop feedback optimization, respectively.

[0098] In a preferred embodiment of the present invention, the water inrush risk assessment also considers spatial distribution characteristics. When multiple low-temperature anomaly areas occur simultaneously, it is necessary to analyze their spatial relationships. If multiple low-temperature anomaly areas are linearly distributed, it indicates the possible existence of interconnected seepage channels, significantly increasing the risk of water inrush; if the low-temperature anomaly areas are scattered, it indicates that the seepage sources are relatively dispersed, and the risk of water inrush is relatively low. Spatial distribution characteristics are quantified by calculating the linear fitting residuals of the centroids of the low-temperature anomaly areas; the smaller the residuals, the higher the linearity and the greater the risk of water inrush.

[0099] Predicting the advance warning also requires considering the orientation of the seepage channel and the attitude of the fault zone. When the seepage channel aligns with the fault zone, groundwater is more likely to flow rapidly along the fault zone, leading to an earlier inrush time. Conversely, when the seepage channel and fault zone orientation form an angle, groundwater needs to overcome greater rock resistance, resulting in a relatively delayed inrush time. This invention corrects the advance warning by analyzing the angle between the principal axis direction of the low-temperature anomaly region and the fault zone orientation. Let the angle be denoted as... The correction factor is The advance warning amount is multiplied by the correction factor to obtain the final forecast value.

[0100] The flood inrush warning signal includes not only the risk level and the lead time for warning, but also recommended emergency response measures. When the flood inrush risk level is high, the system recommends increasing monitoring frequency, preparing drainage equipment, and notifying emergency rescue personnel to stand by. When the flood inrush risk level is extremely high, the system recommends immediately stopping operations at the working face, organizing personnel to evacuate to a safe area, and activating the emergency plan. The recommended emergency response measures are graded according to the lead time for warning; the shorter the lead time, the more urgent the response measures.

[0101] Step S5: Closed-loop feedback optimization steps.

[0102] The closed-loop feedback optimization step receives the water inrush warning signal and warning lead time output from the water inrush precursor feature extraction step, combines this with the actual water inrush occurrence, calculates the warning performance evaluation index, and optimizes the system's threshold parameters accordingly. This step is the key mechanism for achieving continuous performance improvement in this invention. By establishing a closed-loop feedback between the warning result and the actual event, the system can learn from historical data and continuously improve the accuracy and timeliness of warnings.

[0103] Actual water inrush records are collected through manual entry or automatic detection. When a water inrush event occurs, construction personnel record information such as the time of occurrence, location, and intensity of the inrush. Inrush intensity is classified into five levels according to the inrush volume: trace seepage (less than 1 cubic meter per hour), small inrush (1 to 10 cubic meters per hour), medium inrush (10 to 100 cubic meters per hour), large inrush (100 to 1000 cubic meters per hour), and extremely large inrush (greater than 1000 cubic meters per hour).

[0104] The accuracy evaluation index for early warnings reflects the correctness of the system's early warnings. Let the number of early warnings issued within the statistical period be... The number of correct warnings (where water surge actually occurred after a warning) was: The number of false alarms (where no water inrush occurred after the warning) was: The number of missed reports (water inrushes occurring without prior warning) is: The accuracy of the early warning and early warning recall rate They are respectively:

[0105] ,

[0106] ,

[0107] in, The accuracy rate of early warnings represents the proportion of correct early warnings issued. The early warning recall rate represents the proportion of actual water inrush events that received a correct warning. Both indicators together reflect the overall performance of the early warning system.

[0108] The timeliness evaluation index for early warnings reflects the appropriateness of the timing of the early warning signal issuance. For each correct early warning, the time difference between the time of the early warning issuance and the actual time of water inrush is calculated. Ideally, It should fall within the target early warning lead time range (2 to 6 hours). Let the average time difference of correct early warnings within the statistical period be... The evaluation index for the timeliness of early warning Defined as:

[0109] ,

[0110] in, This is an indicator for evaluating the timeliness of early warnings, with a value ranging from 0 to 1. The closer the value is to 1, the more reasonable the timing of the early warning. When the average early warning lead time is less than 2 hours, it indicates that the warning is too late; when the average early warning lead time is greater than 6 hours, it indicates that the warning is too early.

[0111] The threshold parameters are optimized through adaptive adjustment based on early warning performance evaluation indicators. The optimized threshold parameters of this invention include preset temperature deviation values. and preset risk thresholds , , The optimization rules are as follows:

[0112] When the accuracy of the early warning If the accuracy is below the preset accuracy threshold (e.g., 0.8), it indicates that the false alarm rate is too high. In this case, the preset temperature deviation value should be increased. This makes the identification of low-temperature anomaly areas more rigorous and reduces false alarms. The adjusted formula is as follows:

[0113] ,

[0114] in, Adjust the temperature threshold step size, setting it to 0.1 degrees Celsius to 0.5 degrees Celsius.

[0115] When the early warning recall rate If the false negative rate is below the preset recall threshold (e.g., 0.9), it indicates that the false negative rate is too high. In this case, the preset temperature deviation value should be reduced. This improves detection sensitivity. The adjusted formula is:

[0116] ,

[0117] When the early warning timeliness evaluation index This indicates that the warning was issued too early. When this happens, increase the preset risk threshold. and The timing of the early warning trigger will be postponed. The adjusted formula is as follows:

[0118] ,

[0119] in, Adjust the step size for the risk threshold, setting it to 0.02 to 0.05.

[0120] When the early warning timeliness evaluation index This indicates that the warning was issued too late. When this happens, lower the preset risk threshold. and This provides an early warning of when the warning will be triggered. The adjustment formula is similar, but in the opposite direction.

[0121] The optimized threshold parameters are fed back to the infrared temperature field time-series analysis step and the water inrush precursor feature extraction step via a feedback channel, for subsequent identification of low-temperature anomaly areas and determination of water inrush risk level. This closed-loop feedback mechanism enables the system to continuously optimize itself based on actual water inrush conditions, and its early warning performance continuously improves with extended operating time and the accumulation of historical data.

[0122] In a preferred embodiment of the present invention, the closed-loop feedback optimization step further includes online updating of network model parameters. When the number of accumulated correct warning samples reaches a preset threshold, these samples are used as new training data to incrementally train the cross-modal attention fusion network and update the network weight parameters. Incremental training uses a small learning rate (one-tenth of the original training learning rate) and fewer iterations (10 to 20 rounds) to avoid catastrophic forgetting of learned knowledge.

[0123] Through the deep coupling and closed-loop collaboration of the above five steps, this invention achieves automatic identification and early warning of water inrush signs in fault zone tunnels. The multi-source data time-series acquisition step provides high-quality multimodal time-series data; the infrared temperature field time-series analysis step captures the spatiotemporal evolution of the temperature field; the seepage channel multimodal identification step achieves adaptive fusion of infrared and visible light features; the water inrush precursor feature extraction step comprehensively assesses water inrush risk based on multi-dimensional parameters; and the closed-loop feedback optimization step ensures continuous improvement in system performance. The entire method forms a complete closed loop from data acquisition to early warning output and performance optimization, achieving the technical goals of an early warning lead time of 2 to 6 hours and an early warning accuracy rate of over 90%.

[0124] See Figure 2 The present invention also provides an infrared thermal imaging identification system for signs of water inrush in tunnels along fault zones, which corresponds one-to-one with the steps in the above method embodiments. This system includes a multi-source data time-series acquisition module 1, an infrared temperature field time-series analysis module 2, a multi-modal identification module for seepage channels 3, a water inrush precursor feature extraction module 4, and a closed-loop feedback optimization module 5.

[0125] The hardware components of the multi-source data timing acquisition module 1 include an infrared thermal imager, a visible light camera, a temperature sensor array, a water level sensor, and a data synchronization controller. In one embodiment of the invention, the infrared thermal imager is a FLIR A615 uncooled long-wave infrared thermal imager with a temperature resolution of 0.05 degrees Celsius, a spatial resolution of 640×480 pixels, and a spectral response range of 7.5 to 14 micrometers. The visible light camera is a Basler acA1920-40gc industrial camera with a resolution of 1920×1200 pixels and a frame rate of 40fps. The temperature sensor is a PT100 platinum resistance temperature sensor with a measurement accuracy of 0.1 degrees Celsius, arranged at 5 to 10 representative locations on the tunnel face and tunnel wall. The water level sensor is an immersion-type level transmitter with a range of 0 to 50 meters and an accuracy of 0.1% of full scale. The data synchronization controller is designed based on an ARM processor and is responsible for the timing control of the acquisition of each sensor and the data packet transmission.

[0126] The software function implementation method of the multi-source data time-series acquisition module 1 includes the multimodal data registration and temperature compensation correction algorithm described in step S1. Both the registration and correction algorithms are implemented in C++ and run on the CPU of an industrial control computer, with a single frame processing time not exceeding 50 milliseconds.

[0127] The infrared temperature field time-series analysis module 2 implements the temperature field reconstruction, temperature gradient field calculation, low-temperature anomaly region identification, and time-series evolution feature extraction functions described in step S2 of the implementation method embodiment. The core algorithm of this module is implemented in Python, calling NumPy and OpenCV libraries for matrix operations and image processing. Temperature field reconstruction and gradient calculation are executed in parallel on the GPU, with a single frame processing time not exceeding 20 milliseconds. Low-temperature anomaly region identification uses OpenCV's connectedComponents function to perform connected component analysis. The time-series evolution feature extraction maintains a sliding window buffer, storing the analysis results of the most recent 30 frames, and calculates the expansion rate and gradient change rate in real time.

[0128] The implementation method of the multimodal recognition module 3 for seepage channels, as described in step S3, involves a cross-modal attention fusion network. The network model is built using the PyTorch framework. Both the infrared and visible light feature encoders utilize the first four stages of the ResNet-18 backbone network. The cross-modal attention module employs a multi-head attention mechanism with 8 heads. The feature decoder uses transposed convolution for upsampling. The network model runs on an NVIDIA RTX 3080 GPU, with a single-frame inference time of approximately 35 milliseconds when the input image size is 512×512 pixels.

[0129] The water inrush precursor feature extraction module 4 implements the water inrush risk assessment and early warning generation functions described in step S4 of the implementation method embodiment. The risk assessment algorithm of this module is implemented in Python, and historical statistical information of each parameter is maintained for normalization calculation. Early warning signals are pushed to monitoring terminals and mobile devices in real time via the WebSocket protocol. The early warning lead time prediction algorithm dynamically calculates based on the current expansion rate and groundwater level trend, with an output range of 2 to 6 hours.

[0130] The closed-loop feedback optimization module 5 implements the early warning performance evaluation and threshold parameter optimization functions described in step S5 of the implementation method. This module uses an SQLite database to store historical early warning records and water inrush event records, and performs performance evaluation and parameter optimization periodically (e.g., weekly). The optimized threshold parameters are passed to the infrared temperature field time series analysis module 2 and the water inrush precursor feature extraction module 4 via a shared memory mechanism for online updates.

[0131] The system's overall software architecture adopts a modular design, with data transmission between modules via message queues, supporting distributed deployment. The system configuration file allows for flexible adjustment of various parameters, facilitating optimized configuration based on different tunnel geological conditions and construction environments. The system provides a web management interface, supporting functions such as real-time monitoring, historical data querying, parameter configuration, and report generation.

[0132] To verify the technical effectiveness of this invention, a six-month test was conducted at a highway tunnel construction site in a fault zone. During the test, the system collected approximately 500,000 frames of infrared thermal imaging data, detected 128 areas of abnormal low temperatures, issued 15 water inrush warning signals, and 12 actual water inrush events occurred. The test results show that the system's warning accuracy reached 92.3% (12 / 13, 2 false alarms), the warning recall rate reached 100% (12 / 12, no missed alarms), and the average warning lead time was 3.8 hours, meeting the design target of 2 to 6 hours.

[0133] Compared with the method described in CN111899288A, the technical advantages of this invention are reflected in the following aspects: This invention achieves early identification of water inrush signs through time-series analysis, while the comparison document can only detect areas where leakage has already occurred. This invention employs a cross-modal attention fusion mechanism, improving the accuracy of seepage channel identification by approximately 15% compared to the generative adversarial network fusion method of the comparison document. This invention establishes a closed-loop feedback optimization mechanism, with system performance continuously improving over time, while the comparison document uses fixed parameters and cannot adaptively optimize. This invention comprehensively considers multiple dimensions such as the rate of temperature anomaly expansion, the rate of temperature gradient change, seepage intensity, and groundwater level, making the water inrush risk assessment more comprehensive and reliable.

[0134] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for identifying signs of water inrush in tunnels along fault zones using infrared thermal imaging, characterized in that, Includes the following steps: The multi-source data time-series acquisition steps involve acquiring infrared thermal imaging data of the tunnel face and tunnel wall using an infrared thermal imager, acquiring visible light image data using a visible light camera, acquiring ambient temperature data using a temperature sensor, and acquiring groundwater level monitoring data using a water level sensor. Spatial registration is performed between the infrared thermal imaging data and the visible light image data. Temperature compensation correction is then applied to the infrared thermal imaging data based on the ambient temperature data to generate a time-synchronized multimodal data sequence. The infrared temperature field time series analysis steps involve reconstructing the temperature field from the registered and corrected infrared thermal imaging data, calculating the temperature gradient field, identifying low-temperature anomaly regions based on a preset temperature anomaly threshold, and extracting the time series evolution characteristics of the low-temperature anomaly regions, including the expansion rate of the temperature anomaly region and the rate of change of the temperature gradient. The multimodal identification step for seepage channels involves inputting the low-temperature anomaly region and the visible light image data into a cross-modal attention fusion network. Temperature distribution features are extracted through the infrared feature branch, and texture features of the humid area are extracted through the visible light feature branch. The temperature distribution features and the texture features of the humid area are fused through the cross-modal attention mechanism to output the seepage channel identification result. The seepage channel identification result includes the location of the seepage channel and the seepage intensity level. The step of extracting precursor features of water inrush involves calculating water inrush risk assessment indicators based on the temporal evolution features and the seepage channel identification results, combined with the groundwater level monitoring data, determining the water inrush risk level according to the water inrush risk assessment indicators, generating a water inrush early warning signal when the water inrush risk level exceeds a preset risk threshold, and predicting the early warning advance based on the expansion rate of the temperature anomaly zone.

2. The infrared thermal imaging identification method for signs of water inrush in tunnels along fault zones according to claim 1, characterized in that, The identification of low-temperature abnormal regions based on a preset temperature abnormality threshold specifically includes: The temperature value of each pixel in the registered and corrected infrared thermal imaging data is obtained. The average temperature of the tunnel wall region is calculated. The average temperature is subtracted from the product of the standard deviation coefficient and the temperature standard deviation, and then subtracted from the preset temperature deviation value to obtain the dynamic temperature threshold. Pixels with temperature values ​​lower than the dynamic temperature threshold are marked as low-temperature abnormal pixels. Connectivity analysis is performed on the low-temperature abnormal pixels, and the connected components with an area greater than the preset minimum area threshold are determined as low-temperature abnormal regions.

3. The infrared thermal imaging identification method for signs of water inrush in tunnels along fault zones according to claim 1, characterized in that, The calculation of the temperature gradient field includes: The reconstructed temperature field is spatially differentiated along the horizontal and vertical directions to obtain the horizontal gradient component and the vertical gradient component. The gradient magnitude and gradient direction angle are calculated. Pixels with gradient magnitudes greater than a preset gradient threshold are marked as high gradient points. These high gradient points are used to help identify the boundaries of seepage channels.

4. The infrared thermal imaging identification method for signs of water inrush in tunnels along fault zones according to claim 1, characterized in that, The calculation of the expansion rate of the temperature anomaly zone includes: In continuously acquired time-series infrared thermal imaging data, the area of ​​the low-temperature anomaly region at adjacent time points is calculated. The difference between the area at the next time point and the area at the previous time point is divided by the time interval to obtain the instantaneous expansion rate. The instantaneous expansion rate within a preset time window is then averaged to obtain the smoothed expansion rate of the temperature anomaly region.

5. The infrared thermal imaging identification method for signs of water inrush in tunnels along fault zones according to claim 1, characterized in that, The cross-modal attention fusion network includes: An infrared feature encoder employs a coding structure including downsampling convolutional layers and residual connections to extract multi-scale features from the infrared image corresponding to the low-temperature anomaly region, outputting an infrared feature map. A visible light feature encoder uses the same network structure as the infrared feature encoder but with independent network parameters to extract multi-scale features from the visible light image data, outputting a visible light feature map. A cross-modal attention module uses the infrared feature map as a query vector and the visible light feature map as a key vector and value vector, generating a fused feature map through attention calculation. A feature decoder upsamples and reconstructs the fused feature map, outputting the water seepage channel identification result.

6. The infrared thermal imaging identification method for signs of water inrush in tunnels along fault zones according to claim 5, characterized in that, The processing procedure of the cross-modal attention module includes: The infrared feature map is passed through a first linear transformation layer to generate a query matrix. The visible light feature map is passed through a second linear transformation layer and a third linear transformation layer to generate a key matrix and a value matrix, respectively. The product of the query matrix and the transpose of the key matrix is ​​calculated and divided by a scaling factor. The product result is normalized to obtain an attention weight matrix. The attention weight matrix is ​​multiplied by the value matrix to obtain cross-modal fusion features.

7. The infrared thermal imaging identification method for signs of water inrush in tunnels along fault zones according to claim 1, characterized in that, The calculation of the inrush risk assessment indicators includes: The expansion rate of the temperature anomaly zone, the rate of change of the temperature gradient, the infiltration intensity level, and the rate of change of the groundwater level are obtained. The above parameters are normalized respectively. The normalized parameters are weighted and summed to obtain the water inrush risk assessment index. The weight of the weighted summation is determined according to the correlation between each parameter and historical water inrush events.

8. The infrared thermal imaging identification method for signs of water inrush in tunnels along fault zones according to claim 1, characterized in that, The prediction of the advance warning includes: Based on the expansion rate of the temperature anomaly zone and the current area of ​​the low temperature anomaly zone, the time required for the low temperature anomaly zone to reach the preset critical area is predicted. This time is used as the initial estimate of the early warning lead time. The initial estimate is then corrected by combining the changing trend of the groundwater level monitoring data, and the corrected early warning lead time is output. The range of the early warning lead time is two to six hours.

9. The infrared thermal imaging identification method for signs of water inrush in tunnels along fault zones according to claim 1, characterized in that, It also includes a closed-loop feedback optimization step, which records the actual time and intensity of water inrush, compares the actual time of water inrush with the time of water inrush warning signal, calculates the evaluation index of warning accuracy and warning timeliness, adjusts the preset temperature deviation value and the preset risk threshold according to the evaluation index, and feeds back the optimized threshold parameter to the infrared temperature field time series analysis step and the water inrush precursor feature extraction step. The methods for adjusting the threshold parameter in the closed-loop feedback optimization step include: When the accuracy of the early warning is lower than the preset accuracy threshold, the preset temperature deviation value is increased to improve the detection sensitivity; when the early warning timeliness evaluation index indicates that the early warning is too early, the preset risk threshold is increased to reduce false alarms; when the early warning timeliness evaluation index indicates that the early warning is too late, the preset risk threshold is decreased to advance the early warning timing; the threshold adjustment is iteratively updated using a preset step size.

10. An infrared thermal imaging identification system for signs of water inrush in tunnels along fault zones, used to implement the infrared thermal imaging identification method for signs of water inrush in tunnels along fault zones as described in claim 9, characterized in that, include: The multi-source data time-series acquisition module is used to acquire infrared thermal imaging data through an infrared thermal imager, visible light image data through a visible light camera, ambient temperature data through a temperature sensor, and groundwater level monitoring data through a water level sensor. It performs spatial registration and temperature compensation correction on the infrared thermal imaging data and the visible light image data to generate a time-synchronized multimodal data sequence. The infrared temperature field time series analysis module is used to reconstruct the temperature field and calculate the temperature gradient field of the registered and corrected infrared thermal imaging data. Based on the preset temperature anomaly threshold, it identifies low temperature anomaly regions and extracts the time series evolution features, including the expansion rate of the temperature anomaly region and the rate of change of the temperature gradient. The seepage channel multimodal recognition module is used to fuse infrared temperature distribution features and visible light wet area texture features through a cross-modal attention fusion network, and output seepage channel recognition results including the location of the seepage channel and the seepage intensity level. The water inrush precursor feature extraction module is used to calculate water inrush risk assessment indicators based on the time-series evolution features and the seepage channel identification results, determine the water inrush risk level, generate water inrush early warning signals, and predict the early warning lead time. The closed-loop feedback optimization module is used to calculate the evaluation indicators of early warning accuracy and early warning timeliness based on the actual water inrush situation, adjust the preset temperature deviation value and preset risk threshold, and feed back the optimized threshold parameters to the infrared temperature field time series analysis module and the water inrush precursor feature extraction module.

Images