A tunnel blasting quality evaluation method based on multi-source data fusion

Abstract

The present application relates to the technical field of tunnel construction, and particularly relates to a tunnel blasting quality evaluation method based on multi-source data fusion, wherein through a sensor network and an image acquisition device arranged in a tunnel, multi-source data such as blasting vibration waveform, shock wave overpressure, dust concentration time sequence and excavation face image are synchronously collected; the data is preprocessed and time sequence is aligned to form a standardized data sample set; then multi-dimensional feature vectors representing blasting effects are extracted from the data sample set respectively; various features are fused into high-dimensional feature vectors, and a machine learning model is trained by using historical data and quality grade labels evaluated by experts to establish a mapping relationship between features and quality grades, so as to construct an intelligent evaluation model; online automatic evaluation of new blasting events and visualized result output are realized. The present application realizes automation, objectivity and systematization of blasting quality evaluation, and provides an efficient and reliable data-driven solution for tunnel construction quality control.

Description

Technical Field

[0001] This invention relates to the field of tunnel construction technology, and in particular to a method for evaluating the quality of tunnel blasting based on multi-source data fusion. Background Technology

[0002] The evaluation of tunnel blasting quality is a crucial step in tunnel construction, and the accuracy of its results directly impacts construction safety, efficiency, and cost. Traditional evaluation methods primarily rely on technicians conducting on-site inspections at the tunnel face after blasting. They visually observe indicators such as rock face smoothness, half-pore ratio, fragmentation size, and fracture development, combined with vibration monitoring data from individual points, to make subjective, experience-based judgments. This method not only suffers from significant time lag, greatly hindering construction cycle efficiency, but also, due to its heavy reliance on personal experience, makes it difficult to avoid subjectivity and inconsistencies in the evaluation results. Furthermore, it fails to provide objective and comprehensive digital recording and traceability of the entire blasting process.

[0003] With the development of monitoring technology, some existing technologies have attempted to use data from a single or a few sensors for automated evaluation. For example, some rely solely on vibration signals to predict blasting vibration hazards, or solely on image analysis to evaluate rock mass blockiness. However, these methods often start from a single perspective or physical phenomenon, failing to comprehensively consider the intrinsic relationships between multiple factors such as the vibration effects, air shock wave effects, dust diffusion effects, and changes in the apparent morphology of the rock mass generated by blasting. This results in a single evaluation dimension, making it difficult to comprehensively and accurately reflect the overall condition of blasting quality, and significantly limiting the reliability and universality of their evaluation conclusions.

[0004] In recent years, big data and artificial intelligence technologies have offered new solutions to these challenges. Theoretically, by fusing multi-source heterogeneous data from various types of sensors throughout the blasting process and utilizing machine learning models to uncover the deep, nonlinear mapping relationship between this data and blasting quality, it is hoped that an objective, efficient, and comprehensive intelligent evaluation method can be constructed. However, effectively achieving simultaneous acquisition, standardized preprocessing, and feature layer fusion of multi-source data, and constructing a high-precision evaluation model accordingly, remains a pressing technical challenge in this field. Summary of the Invention

[0005] According to a first aspect of the present invention, the present invention claims protection for a method for evaluating the quality of tunnel blasting based on multi-source data fusion, comprising:

[0006] S1, within a set distance behind the tunnel blasting excavation face, a monitoring network consisting of multiple types of sensors is simultaneously deployed and triggered. After the blasting operation is completed, the timing signal data stream generated by all sensors and the digital image of the excavation face are recorded within a preset time window starting from the moment of blasting.

[0007] S2, the time-series signal data stream and the digital image of the excavation face are cleaned and standardized, and the time-series signal data stream is aligned with the image acquisition time, and associated with the same blasting event to configure event identifiers to form standardized multi-source data packets;

[0008] S3, calculate the characteristic parameters that quantitatively characterize the blasting effect from the standardized multi-source data packets respectively;

[0009] S4, combine the feature parameters into a comprehensive feature vector in a predetermined order, configure the historical database to include the comprehensive feature vector and the corresponding quality grade label manually evaluated based on the actual on-site investigation results after blasting and with reference to the blasting design target;

[0010] S5. Use the historical database to train a machine learning model. During training, adjust the internal parameters of the model iteratively to learn the complex mapping relationship between the comprehensive feature vector and the quality level label, and obtain a trained intelligent evaluation model for blasting quality.

[0011] S6. For a new blasting operation, repeat steps S1 to S3 to obtain a new comprehensive feature vector and input it into the trained intelligent blasting quality evaluation model. The model will automatically output the predicted quality level and a visual report.

[0012] Furthermore, step S1 also includes:

[0013] The monitoring network includes at least a three-component vibration sensor for measuring rock mass vibration acceleration, an overpressure sensor for measuring air shock wave pressure, a laser dust sensor for measuring the concentration of suspended particulate matter in the air, and a high-resolution explosion-proof digital camera for collecting the surface condition of the excavation face after blasting.

[0014] The three-component vibration sensor is fixed to the tunnel sidewall or the rock mass behind the working face by plaster coupling.

[0015] The overpressure sensor and the laser dust sensor are installed on the sidewall support of the same section of the tunnel, with the sensor head facing the excavation face;

[0016] After the explosion, the high-resolution explosion-proof digital camera was moved to a safe location and then remotely operated by staff to acquire images.

[0017] Furthermore, step S2 also includes:

[0018] S2.3a, call the camera lens distortion parameter matrix obtained in advance through the checkerboard calibration method, perform mathematical transformation on the coordinates of each pixel point of the image, eliminate the radial and tangential distortion caused by the inherent characteristics of the lens, and obtain an image with a geometry closer to the real scene;

[0019] S2.3b: In the image after distortion correction, automatically identify multiple calibration targets with specific shapes and known physical dimensions that are pre-pasted on the excavation surface and are known to be specific shapes. Based on the correspondence with the real-world coordinates, calculate the perspective transformation matrix, resample the entire image, and convert the image into an orthographic projection plane.

[0020] S2.3c, On the orthophoto plane, an image segmentation algorithm based on pixel color and texture features is used to divide the area into multiple regions and extract the region of interest. The brightness value histogram of all pixels in the region of interest is calculated, and the overall brightness and contrast are adjusted to a predefined standard level.

[0021] Furthermore, step S3 also includes:

[0022] S3.4a, On the rock mass area image after brightness normalization, the edge detection algorithm is used to identify the obvious contour lines and uneven boundaries of the rock mass surface, and the linear structure recognition algorithm is used to detect line segments that are close to straight lines from the edges. The total length, uncertainty of the direction angle distribution, and average vertical distance deviation between all detected line segments in a unit area and the designed excavation contour line are statistically analyzed, and the statistical quantities are combined into a flatness feature.

[0023] S3.4b, apply a region segmentation algorithm based on image grayscale and texture features to merge connected pixel regions with similar features, segment the rock mass into individual rock block regions and calculate the pixel area, convert it into the real physical area or equivalent diameter according to the image scale, statistically analyze the distribution of all identified equivalent diameters, and record the median grain size and the uniformity index of the grain size distribution.

[0024] S3.4c uses an directionally adjustable filter bank to enhance the image, highlighting linear features in different directions. The enhanced image is then binarized through thresholding to obtain a potential gap pixel network.

[0025] S3.4d performs morphological operations on the binarized image to connect broken crack segments and remove small noise points, and calculates the total pixel length of the final crack network per unit area, the complexity connectivity index of the crack network, and the average pixel width of the cracks.

[0026] Furthermore, in step S5, the machine learning model is a gradient boosting decision tree model;

[0027] The training process gradually improves prediction accuracy by fitting the prediction residuals of previous models to each decision tree sequentially. The combination of hyperparameters of the model, including the maximum depth of the tree, the learning rate, and the number of subtrees, is determined by cross-validation combined with an automated parameter search strategy.

[0028] Furthermore, when training the gradient boosting decision tree model, the historical database is randomly divided into a subset for direct training, a subset for real-time performance verification and early stopping during training to prevent overfitting, and a test subset for evaluating the model's generalization ability.

[0029] Furthermore, in step S4, after constructing the comprehensive feature vector and before inputting it into the model for training, each feature parameter in the vector is numerically scaled.

[0030] The feature value is mapped to a range between zero and one by subtracting the minimum value of historical samples from the feature value of the comprehensive feature vector and then dividing by the range of historical samples.

[0031] Furthermore, in step S5, the hyperparameters of the machine learning model are optimized and selected using the Bayesian optimization method;

[0032] We construct a probabilistic surrogate model to fit the relationship between the hyperparameter combination and the model validation performance, and find the hyperparameter combination with the best validation performance through iteration.

[0033] Furthermore, the visualization report described in step S6 includes parallel bar charts or radar charts;

[0034] Each chart represents a key feature parameter, indicating the specific value of that parameter in this blasting and a reference range representing the numerical distribution of the parameter in the historical database of blasting samples of different grades.

[0035] Furthermore, the method also includes a model update loop step S7:

[0036] The system periodically checks newly accumulated blasting data samples whose quality levels have been confirmed by on-site engineers.

[0037] The blasting data samples are added to the historical database, and the intelligent evaluation model for blasting quality is retrained or incrementally learned to update the model parameters to adapt to changes in geological conditions or blasting parameters during tunnel excavation.

[0038] This invention relates to the field of tunnel construction technology, specifically a method for evaluating the quality of tunnel blasting based on multi-source data fusion. The method utilizes a sensor network and image acquisition equipment deployed within the tunnel to simultaneously collect multi-source data, including blasting vibration waveforms, shock wave overpressure, dust concentration time series, and images of the excavation face. The data is preprocessed and aligned with time series to form a standardized data sample set. Multi-dimensional feature vectors characterizing the blasting effect are then extracted from these samples. These features are fused into a high-dimensional feature vector, and a machine learning model is trained using historical data and expert-assessed quality grade labels to establish a mapping relationship between features and quality grades, thus constructing an intelligent evaluation model. This enables online automatic evaluation of new blasting events and visualized result output. This invention achieves automation, objectivity, and systematization of blasting quality evaluation, providing an efficient and reliable data-driven solution for tunnel construction quality control. Attached Figure Description

[0039] Figure 1 The flowchart illustrates a tunnel blasting quality evaluation method based on multi-source data fusion, which is claimed in an embodiment of the present invention.

[0040] Figure 2 This is a second flowchart of a tunnel blasting quality evaluation method based on multi-source data fusion, which is claimed in an embodiment of the present invention.

[0041] Figure 3 The third flowchart is a method for evaluating the quality of tunnel blasting based on multi-source data fusion, which is claimed in an embodiment of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0043] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0044] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0045] According to the first embodiment of the present invention, referring to Figure 1 This invention claims protection for a method for evaluating the quality of tunnel blasting based on multi-source data fusion, comprising:

[0046] S1, within a set distance behind the tunnel blasting excavation face, a monitoring network consisting of multiple types of sensors is simultaneously deployed and triggered. After the blasting operation is completed, the timing signal data stream generated by all sensors and the digital image of the excavation face are recorded within a preset time window starting from the moment of blasting.

[0047] S2, the time-series signal data stream and the digital image of the excavation face are cleaned and standardized, and the time-series signal data stream is aligned with the image acquisition time, and associated with the same blasting event to configure event identifiers to form standardized multi-source data packets;

[0048] S3, calculate the characteristic parameters that quantitatively characterize the blasting effect from the standardized multi-source data packets respectively;

[0049] S4, combine the feature parameters into a comprehensive feature vector in a predetermined order, configure the historical database to include the comprehensive feature vector and the corresponding quality grade label manually evaluated based on the actual on-site investigation results after blasting and with reference to the blasting design target;

[0050] S5. Use the historical database to train a machine learning model. During training, adjust the internal parameters of the model iteratively to learn the complex mapping relationship between the comprehensive feature vector and the quality level label, and obtain a trained intelligent evaluation model for blasting quality.

[0051] S6. For a new blasting operation, repeat steps S1 to S3 to obtain a new comprehensive feature vector and input it into the trained intelligent blasting quality evaluation model. The model will automatically output the predicted quality level and a visual report.

[0052] This embodiment details the specific implementation process of the method in a typical tunnel engineering project. The implementation process strictly follows each step, aiming to demonstrate the feasibility and completeness of the method through detailed technical specifications.

[0053] A distributed monitoring system is constructed in a stable area at a specific distance behind the tunnel blasting excavation face. This system consists of various types of monitoring equipment: three-component vibration sensors are installed on the tunnel sidewalls and arch, with the rock surface ground smoothed before installation. Rapid-curing plaster is used to tightly couple the sensor housing to the rock mass, ensuring high-fidelity transmission of vibration signals. Overpressure sensors are fixed to rigid mounting rods, with the sensor diaphragm facing the center of the excavation face. The mounting rod base is anchored in the rock mass using expansion bolts. Laser dust sensors are installed in a dedicated protective box with a sampling port aligned with the tunnel centerline to avoid wall effects. A high-resolution explosion-proof digital camera is mounted on a heavy-duty pan-tilt unit capable of omnidirectional rotation, with the pan-tilt base fixed to a pre-embedded component in the tunnel floor at a specific distance from the excavation face. Upon issuance of a blasting command, all sensors simultaneously begin high-speed continuous data recording. After blasting is completed, ventilation and dust removal are performed, and site safety is confirmed, technicians remotely control the pan-tilt unit and camera to capture high-definition digital images of the excavation face from multiple preset angles, ensuring coverage of the entire excavation area. All data recorded by the monitoring equipment is accompanied by a high-precision timestamp obtained from the unified time synchronization system and transmitted to an external data server for storage via a wired network.

[0054] After receiving the raw data, the server initiates an automatic preprocessing procedure: For vibration and overpressure data streams, a data quality check is first performed, automatically identifying and marking abnormal data segments where the signal amplitude exceeds the physically reasonable range or where abrupt plateaus occur. Abnormal segments are replaced using linear interpolation of the preceding and following valid data. Subsequently, the data passes through a pre-set digital low-pass filter at a cutoff frequency to suppress high-frequency electronic noise while preserving the main physical signal frequency bands generated by the explosion.

[0055] For the dust concentration data sequence, check for signal saturation due to excessively high instantaneous concentration or signal loss due to dust obstructing the laser beam. For identified data loss periods, use a linear interpolation algorithm to fill in the gaps based on the data change trends of the preceding and following periods. Apply a moving average filter to the entire sequence to smooth out random fluctuations and better reflect the overall trend of dust concentration changes.

[0056] For the acquired multiple digital images, the preprocessing process includes the following steps: First, the camera lens calibration parameters embedded in the image file are read, and these parameters are applied to perform mathematical transformations on the image to correct the image edge distortion caused by lens spherical distortion and assembly errors. Then, the image processing algorithm automatically identifies the pre-pasted physical calibration targets with known precise size and shape in the image. Based on the correspondence between the two-dimensional pixel coordinates of multiple targets in the image and their three-dimensional design coordinates in the tunnel space, a perspective transformation matrix is ​​calculated. This matrix is ​​then applied to resample the image to generate a planar image that conforms to the orthographic projection rules and has a uniform scale. Finally, considering the differences in image brightness and color, the algorithm segments the main rock mass region in the image, calculates the brightness statistics of the pixels in this region, and normalizes it to a standard brightness range by adjusting the image gamma value and contrast.

[0057] After all data has been preprocessed individually, the system uses the trigger signal time of the blasting network detonator as the absolute time zero point, and precisely aligns the time sequence signal data streams of all sensors with the image acquisition time to ensure that all data segments strictly correspond to the same blasting event. Finally, the data is packaged into a standardized data packet containing multimodal data.

[0058] From the aligned and standardized data packets, the system automatically executes a feature extraction program. For vibration feature extraction, it calculates the three-component vibration velocity time-history curves separately. The time-domain signal is converted into a frequency-domain energy distribution using a Fast Fourier Transform, identifying the frequency band with the most concentrated energy and taking its center frequency as the dominant frequency. The entire time-history curve is scanned to find the maximum absolute amplitude among the three components, which is then converted into peak particle velocity based on the sensor calibration coefficient. The time interval from when the signal first continuously exceeds the background noise threshold until the signal finally decays and remains below that threshold is calculated as the vibration duration.

[0059] For shock wave feature extraction, the maximum positive pressure value is found by scanning the overpressure time history curve; the duration of this positive pressure effect is calculated from the first time the pressure curve exceeds the ambient air pressure to the last time it falls back to the ambient air pressure.

[0060] When extracting dust features, the global maximum value of the dust concentration time history curve is directly found; the time taken from the point where the concentration begins to rise significantly after the explosion to the peak point, and the time taken from the peak point to a point where the concentration drops to a certain percentage of the peak point are calculated to comprehensively characterize the dust diffusion rate.

[0061] During image feature extraction, the following analysis is performed on the preprocessed orthophoto-projected excavation face image:

[0062] For fine segmentation of the rock mass area, an image segmentation algorithm based on color clustering and texture analysis is employed to classify pixels in the image into different categories such as fresh rock surface, existing rock surface, shadow, water accumulation, metal equipment, and abandoned drill pipes. By setting an area threshold and performing morphological operations, the largest continuous fresh rock surface area is extracted as the region of interest for subsequent analysis.

[0063] When calculating the flatness statistics, an edge detection operator is used within the region of interest to identify obvious light and dark transition boundaries on the rock surface; a line detection algorithm is then applied to find approximately linear segments from these boundaries; the unit density of these line segments, the angular distribution dispersion of the line segment directions, and the degree of deviation of all line segment points from the average distance of the designed excavation outline are statistically analyzed.

[0064] Block size distribution calculation: A gradient edge-based region segmentation algorithm is used to close the boundary lines with obvious gradient changes in the image, forming independent blocks; the pixel area of ​​each block is calculated and converted into actual square meters or equivalent diameter according to the image scale; the equivalent diameter of all identified blocks is counted to form a distribution histogram, and the median particle size and particle size distribution uniformity index are extracted from it.

[0065] Quantification of crack development degree: Multi-directional filter bank is applied to the image for convolution operation to enhance the linear features of different directions; the enhanced image is converted into a binary image through adaptive thresholding to highlight potential cracks; morphological closing operation is performed on the binary image to connect small breakpoints, and opening operation is performed to remove isolated noise points; for the finally obtained connected crack network, its total length, the topological connection complexity of the network (such as the ratio of the number of branch points to the number of endpoints) are calculated, and the average width is estimated.

[0066] All feature parameters extracted in step S3, such as vibration, shock wave, dust, image smoothness, image block size, and image cracks, are concatenated into a one-dimensional comprehensive feature vector in a predefined order. A database containing a large number of historical blasting samples is prepared. Each sample contains such a comprehensive feature vector and its corresponding authoritative quality grade label, given by multiple senior blasting engineers based on on-site measurements of over-excavation / under-excavation, half-hole ratio, fragment size measurement, crack mapping, and comprehensive judgment. A gradient boosting decision tree framework is used as the machine learning model.

[0067] The model training process builds a series of decision trees in an iterative manner, with each new tree dedicated to correcting the residuals predicted by all previous tree combinations;

[0068] The parameters of each tree are optimized by minimizing the loss function between the predicted label and the true rank label;

[0069] The key structural hyperparameters of the model are determined by searching on a separate validation dataset using an automated optimization strategy to find the optimal configuration, and finally trained to generate a stable intelligent evaluation model for blasting quality.

[0070] For any new blast, the system automatically executes steps S1 to S3 to generate a new comprehensive feature vector. This vector is input into a pre-trained evaluation model, which internally calculates and outputs a discrete quality level prediction (e.g., "Good"). Simultaneously, the system automatically invokes the report generation module. This module extracts the statistical distribution range of historical "Excellent" samples across various features from the database and plots the feature values ​​of the current blast in bar chart form. Each bar chart has a shaded background clearly indicating the common numerical range of historical "Excellent" samples for that feature, generating a visual comparative report for rapid on-site decision-making.

[0071] Furthermore, step S1 also includes:

[0072] The monitoring network includes at least a three-component vibration sensor for measuring rock mass vibration acceleration, an overpressure sensor for measuring air shock wave pressure, a laser dust sensor for measuring the concentration of suspended particulate matter in the air, and a high-resolution explosion-proof digital camera for collecting the surface condition of the excavation face after blasting.

[0073] The three-component vibration sensor is fixed to the tunnel sidewall or the rock mass behind the working face by plaster coupling.

[0074] The overpressure sensor and the laser dust sensor are installed on the sidewall support of the same section of the tunnel, with the sensor head facing the excavation face;

[0075] After the explosion, the high-resolution explosion-proof digital camera was moved to a safe location and then remotely operated by staff to acquire images.

[0076] Furthermore, referring to Figure 2 Step S2 also includes:

[0077] S2.3a, call the camera lens distortion parameter matrix obtained in advance through the checkerboard calibration method, perform mathematical transformation on the coordinates of each pixel point of the image, eliminate the radial and tangential distortion caused by the inherent characteristics of the lens, and obtain an image with a geometry closer to the real scene;

[0078] S2.3b: In the image after distortion correction, automatically identify multiple calibration targets with specific shapes and known physical dimensions that are pre-pasted on the excavation surface and are known to be specific shapes. Based on the correspondence with the real-world coordinates, calculate the perspective transformation matrix, resample the entire image, and convert the image into an orthographic projection plane.

[0079] S2.3c, On the orthophoto plane, an image segmentation algorithm based on pixel color and texture features is used to divide the area into multiple regions and extract the region of interest. The brightness value histogram of all pixels in the region of interest is calculated, and the overall brightness and contrast are adjusted to a predefined standard level.

[0080] In this embodiment, it further includes:

[0081] For the timing signals output by the vibration sensor and the overpressure sensor, a discrimination method based on standard thresholds is used to identify and eliminate abnormal jump points caused by equipment interference, and a specified electronic filter is used for smoothing and noise reduction.

[0082] For the concentration time series output by the dust sensor, a method based on the weighted average of the data before and after the time series is used to fill in the missing values ​​caused by the brief interruption of the signal, and the entire series is filtered to eliminate high-frequency noise.

[0083] For images acquired by digital cameras, distortion correction is performed based on the inherent optical parameters of the cameras. By identifying reference markers with known physical dimensions that are pre-fixed on the excavation surface, the images are transformed into orthographic projection views that eliminate perspective effects. The rock mass areas in the images are then normalized based on brightness statistics.

[0084] After individual preprocessing is completed, the absolute time point at which the blasting command is issued is used as the unified time reference origin.

[0085] For optical distortion correction, the system incorporates lens distortion parameters pre-calibrated for each camera in a laboratory environment using a high-precision checkerboard calibration plate. During processing, for each pixel in the image, based on its original coordinates and the lens distortion model, including radial and tangential distortion coefficients, its corrected coordinates under an ideal distortion-free lens are calculated. The pixel value at that location is then calculated using bilinear interpolation to generate an image free of optical distortion.

[0086] During orthographic projection transformation, in the distortion-corrected image, the image recognition algorithm automatically detects multiple circular or square calibration targets of known shape and size. Based on the two-dimensional pixel coordinates of the targets in the image and their true coordinates in the three-dimensional space of the tunnel design, a precise projection transformation matrix is ​​calculated using methods such as direct linear transformation. This matrix is ​​then applied to perform inverse perspective transformation and resampling on the entire image, generating a uniform-scale planar orthographic image facing the excavation face, where each pixel represents a fixed dimension on the ground.

[0087] For rock mass region segmentation and brightness normalization, an unsupervised clustering algorithm such as K-means is used on the orthophoto image to initially segment the image into different categories based on pixel color and local texture features. Combining predefined rules, such as spatial location, region area, and color HSV range, the main region representing the freshly excavated rock surface is selected from these categories, and its accurate binary mask image is generated. Using this mask, image regions containing only the main rock mass are extracted, and the brightness value distribution histogram of the pixels in this region is analyzed. Through nonlinear grayscale transformation gamma correction and histogram matching techniques, the brightness and contrast of this region's image are adjusted to a level consistent with a preset standard template image, effectively eliminating image quality differences caused by different lighting conditions during shooting.

[0088] Furthermore, referring to Figure 3 Step S3 also includes:

[0089] S3.4a, On the rock mass area image after brightness normalization, the edge detection algorithm is used to identify the obvious contour lines and uneven boundaries of the rock mass surface, and the linear structure recognition algorithm is used to detect line segments that are close to straight lines from the edges. The total length, uncertainty of the direction angle distribution, and average vertical distance deviation between all detected line segments in a unit area and the designed excavation contour line are statistically analyzed, and the statistical quantities are combined into a flatness feature.

[0090] S3.4b, apply a region segmentation algorithm based on image grayscale and texture features to merge connected pixel regions with similar features, segment the rock mass into individual rock block regions and calculate the pixel area, convert it into the real physical area or equivalent diameter according to the image scale, statistically analyze the distribution of all identified equivalent diameters, and record the median grain size and the uniformity index of the grain size distribution.

[0091] S3.4c uses an directionally adjustable filter bank to enhance the image, highlighting linear features in different directions. The enhanced image is then binarized through thresholding to obtain a potential gap pixel network.

[0092] S3.4d performs morphological operations on the binarized image to connect broken crack segments and remove small noise points, and calculates the total pixel length of the final crack network per unit area, the complexity connectivity index of the crack network, and the average pixel width of the cracks.

[0093] In this embodiment, the main oscillation frequency components of the waveform in each coordinate direction are identified from the preprocessed vibration waveform data, the maximum amplitude value of the waveform is measured and converted into peak particle velocity, and the total time taken for the vibration signal to decay from the beginning to the background noise level is calculated.

[0094] From the preprocessed shock wave data, identify the maximum pressure value of the waveform and calculate the duration during which the pressure value exceeds the ambient atmospheric pressure;

[0095] From the pre-processed dust concentration data, identify the maximum concentration and calculate the time rate during which the concentration decreases from the rising phase to the maximum value and then to the falling phase, thus decaying back to the preset peak ratio.

[0096] From the preprocessed rock mass image, the region representing the rock mass surface is segmented from other non-interested regions. Statistics used to describe the degree of surface unevenness are calculated. The outline of individual rock blocks is identified through image segmentation and the size distribution is statistically analyzed. The development degree of visible fractures in the image is detected and quantified.

[0097] For the calculation of flatness statistics, a Canny edge detector is applied to the preprocessed rock mass image. Strong and weak edges are identified by setting high and low thresholds, and weak edges are connected to form a complete rock surface contour boundary. Subsequently, a probabilistic Hough transform is used to detect line segments from the edge map. Minimum line segment length and maximum line segment gap parameters are set to filter out short and meaningless line segments. For all detected valid line segments, the ratio of their total length to the area of ​​the rock mass region is calculated to obtain the line density. The standard deviation or information entropy of the directional angles of all line segments is calculated to quantify the directional disorder. Each line segment is projected onto the design contour line, and the average vertical distance from each point in the line segment to the design line is calculated as the average deviation.

[0098] The rock mass image was processed using a marker-controlled watershed segmentation algorithm to calculate the rock block size distribution. First, the gradient magnitude map of the image was calculated, with high gradients representing rock block boundaries. To prevent over-segmentation, morphological operations and distance transformations were used to generate reliable "marker" images, corresponding to the internal regions of each rock block. The marked images and gradient images were input into the watershed algorithm, which ultimately segmented the image into independent, closed regions, each representing a rock block. The pixel area of ​​each segmented region was calculated and converted to an actual area based on the image resolution. The area distribution of all rock blocks was statistically analyzed, and their cumulative distribution curves were calculated. The equivalent diameter corresponding to the median area was then obtained, and a uniformity coefficient reflecting the distribution width was calculated.

[0099] To quantify the degree of fracture development, a Steerable filter is used to filter the image in different directions. The direction and amplitude of the strongest response are recorded to enhance fractures with different orientations. Thresholding segmentation is performed on the filtered image to generate a preliminary binary fracture map. Morphological closing operations are then performed on this binary map, using a structuring element of a specific size for dilation followed by erosion to bridge narrow fracture gaps; subsequently, an opening operation is performed, first eroding and then dilation, to eliminate noise points smaller than the structuring element. Connectivity analysis is then performed on the final processed binary map, calculating the total pixel length of fractures in all connected regions, using the Euler number (the difference between the number of holes and the number of objects) to describe the topological complexity of the fracture network, and calculating the average width of all fracture pixels.

[0100] Furthermore, in step S5, the machine learning model is a gradient boosting decision tree model;

[0101] The training process gradually improves prediction accuracy by fitting the prediction residuals of previous models to each decision tree sequentially. The combination of hyperparameters of the model, including the maximum depth of the tree, the learning rate, and the number of subtrees, is determined by cross-validation combined with an automated parameter search strategy.

[0102] In this embodiment, the gradient boosting decision tree model is specifically selected for model S4 construction. Its training is a sequential, additive process. During initialization, the model uses a simple constant value as the initial prediction. Then, decision trees are built iteratively: in each iteration, the gradient direction of the prediction residuals for each training sample is calculated by summing the predictions of all currently built trees. A new decision tree is then constructed, its learning objective being to fit these residuals. This new tree maximizes the reduction of the sum of squared residuals by recursively selecting features and split points. After the new tree is built, it is added to the current model with a small learning rate. Through numerous such iterations, the model gradually improves its predictive ability.

[0103] Furthermore, when training the gradient boosting decision tree model, the historical database is randomly divided into a subset for direct training, a subset for real-time performance verification and early stopping during training to prevent overfitting, and a test subset for evaluating the model's generalization ability.

[0104] In this embodiment, a strict data partitioning strategy is employed to ensure the model's generalization ability when training the gradient boosting decision tree model. The historical database is randomly shuffled and divided into three mutually exclusive subsets: a training subset, used for directly training the model (i.e., calculating gradients and constructing each decision tree); a validation subset, which does not participate in the training itself but is used to calculate the current prediction error after adding a new tree in each iteration. The training process is terminated early when the validation error stops decreasing or even begins to increase in consecutive iterations to prevent the model from overfitting the training data; and a test subset, used for the final, independent evaluation of the model's generalization performance after the entire training and validation process is complete. This subset does not participate in any decisions regarding model structure selection, hyperparameter tuning, or early termination.

[0105] Furthermore, in step S4, after constructing the comprehensive feature vector and before inputting it into the model for training, each feature parameter in the vector is numerically scaled.

[0106] The feature value is mapped to a range between zero and one by subtracting the minimum value of historical samples from the feature value of the comprehensive feature vector and then dividing by the range of historical samples.

[0107] In this embodiment, before inputting the concatenated high-dimensional feature vector into the gradient boosting decision tree model for training, feature scaling preprocessing is performed. For each feature dimension, the maximum and minimum values ​​of that feature are calculated separately from all samples in the training subset. Then, for any sample, including training, validation, and test samples, the following linear transformation is applied to the feature value: Scaled value = (Original value - Minimum value) / (Maximum value - Minimum value). After this processing, all values ​​of each feature are mapped to zero and one. This step ensures that features with different physical dimensions and numerical ranges have equal importance in model training, which helps improve the model's convergence speed and final performance.

[0108] Furthermore, in step S5, the hyperparameters of the machine learning model are optimized and selected using the Bayesian optimization method;

[0109] We construct a probabilistic surrogate model to fit the relationship between the hyperparameter combination and the model validation performance, and find the hyperparameter combination with the best validation performance through iteration.

[0110] In this embodiment, a Bayesian optimization framework is used to automatically find the optimal hyperparameter combination for the gradient boosting decision tree model. This optimization process treats hyperparameter search as a black-box optimization problem. It constructs a probabilistic surrogate model, such as a Gaussian process, to approximate the unknown functional relationship between the hyperparameter combination and the model's accuracy performance on the validation set. Based on this surrogate model and a defined acquisition function, the optimization algorithm intelligently selects the next set of hyperparameter points to be evaluated. After evaluating the performance of each point, the surrogate model is updated with new data. This process is iterated until a predetermined number of evaluations is reached. Finally, the hyperparameter combination that performs best on the validation set throughout the entire optimization process is selected as the configuration of the final model.

[0111] Furthermore, the visualization report described in step S6 includes parallel bar charts or radar charts;

[0112] Each chart represents a key feature parameter, indicating the specific value of that parameter in this blasting and a reference range representing the numerical distribution of the parameter in the historical database of blasting samples of different grades.

[0113] In this embodiment, the generated visualization report creates a horizontal bar chart for each selected key feature parameter, such as peak particle velocity. The length of the bar chart represents the actual measured value of that feature in this blast. Against the background of the bar chart's coordinate axis, a softly colored band is plotted. The left and right boundaries of this band correspond to a statistical range for that feature across all samples in the historical database that were rated "excellent" by experts, for example, from the lower quartile to the upper quartile. This visualization method allows report users to intuitively compare the relative positions and differences between the various indicators of this blast and historical "excellent" levels.

[0114] Furthermore, the method also includes a model update loop step S7:

[0115] The system periodically checks newly accumulated blasting data samples whose quality levels have been confirmed by on-site engineers.

[0116] The blasting data samples are added to the historical database, and the intelligent evaluation model for blasting quality is retrained or incrementally learned to update the model parameters to adapt to changes in geological conditions or blasting parameters during tunnel excavation.

[0117] This embodiment also includes a model update and maintenance mechanism. A trigger condition is set; for example, after a certain number of blasting cycles are successfully completed and entered, when the condition is met, the system automatically checks the data of these new blasting cycles: whether the multi-source data is complete, whether the preprocessing was successful, and most importantly, whether the blasting quality level has been confirmed and entered by the on-site engineer based on the field acceptance results. This batch of new, confirmed sample data is added to the original historical database, forming an expanded training dataset. Subsequently, the system can automatically or manually trigger the model retraining process using the completely new expanded dataset, re-executing the entire process from feature normalization to model training and validation, thereby generating an updated version of the blasting quality evaluation model. After the new model is validated and approved, it can be deployed to the production environment to gradually replace the old model, enabling the evaluation system to continuously adapt to changes in geological conditions ahead of the tunnel and the impact of adjustments to blasting parameters.

[0118] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0119] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0120] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A method for evaluating the quality of tunnel blasting based on multi-source data fusion, characterized in that, include: S1, within a set distance behind the tunnel blasting excavation face, a monitoring network consisting of multiple types of sensors is simultaneously deployed and triggered. After the blasting operation is completed, the timing signal data stream generated by all sensors and the digital image of the excavation face are recorded within a preset time window starting from the moment of blasting. S2, the time-series signal data stream and the digital image of the excavation face are cleaned and standardized, and the time-series signal data stream is aligned with the image acquisition time, and associated with the same blasting event to configure event identifiers to form standardized multi-source data packets; S3, calculate the characteristic parameters that quantitatively characterize the blasting effect from the standardized multi-source data packets; the characteristic parameters that characterize the blasting effect also include vibration characteristics, shock wave characteristics and dust characteristics; S4, combine the feature parameters into a comprehensive feature vector in a predetermined order, configure the historical database to include the comprehensive feature vector and the corresponding quality grade label manually evaluated based on the actual on-site investigation results after blasting and with reference to the blasting design target; S5. Use the historical database to train a machine learning model. During training, adjust the internal parameters of the model iteratively to learn the complex mapping relationship between the comprehensive feature vector and the quality level label, and obtain a trained intelligent evaluation model for blasting quality. S6. For a new blasting operation, repeat steps S1 to S3 to obtain a new comprehensive feature vector and input it into the trained intelligent evaluation model for blasting quality. The predicted quality level result will be automatically output and a visualization report will be output. Step S3 also includes: S3.4a, On the rock mass area image after brightness normalization, the edge detection algorithm is used to identify the obvious contour lines and uneven boundaries of the rock mass surface, and the linear structure recognition algorithm is used to detect line segments that are close to straight lines from the edges. The total length, uncertainty of the direction angle distribution, and average vertical distance deviation between all detected line segments in a unit area and the designed excavation contour line are statistically analyzed, and the statistical quantities are combined into a flatness feature. S3.4b, apply a region segmentation algorithm based on image grayscale and texture features to merge connected pixel regions with similar features, segment the rock mass into individual rock block regions and calculate the pixel area, convert it into the real physical area or equivalent diameter according to the image scale, statistically analyze the distribution of all identified equivalent diameters, and record the median grain size and the uniformity index of the grain size distribution. S3.4c uses an directionally adjustable filter bank to enhance the image, highlighting linear features in different directions. The enhanced image is then binarized through thresholding to obtain a potential gap pixel network. S3.4d performs morphological operations on the binarized image to connect broken crack segments and remove small noise points, and calculates the total pixel length of the final crack network per unit area, the complexity connectivity index of the crack network, and the average pixel width of the cracks.

2. The tunnel blasting quality evaluation method based on multi-source data fusion according to claim 1, characterized in that, Step S1 also includes: The monitoring network includes at least a three-component vibration sensor for measuring rock mass vibration acceleration, an overpressure sensor for measuring air shock wave pressure, a laser dust sensor for measuring the concentration of suspended particulate matter in the air, and a high-resolution explosion-proof digital camera for collecting the surface condition of the excavation face after blasting. The three-component vibration sensor is fixed to the tunnel sidewall or the rock mass behind the working face by plaster coupling. The overpressure sensor and the laser dust sensor are installed on the sidewall support of the same section of the tunnel, with the sensor head facing the excavation face; After the explosion, the high-resolution explosion-proof digital camera was moved to a safe location and then remotely operated by staff to acquire images.

3. The tunnel blasting quality evaluation method based on multi-source data fusion according to claim 1, characterized in that, Step S2 also includes: S2.3a, call the camera lens distortion parameter matrix obtained in advance through the checkerboard calibration method, perform mathematical transformation on the coordinates of each pixel point of the image, eliminate the radial and tangential distortion caused by the inherent characteristics of the lens, and obtain an image with a geometry closer to the real scene; S2.3b: In the image after distortion correction, automatically identify multiple calibration targets with specific shapes and known physical dimensions that are pre-pasted on the excavation surface and are known to be specific shapes. Based on the correspondence with the real-world coordinates, calculate the perspective transformation matrix, resample the entire image, and convert the image into an orthographic projection plane. S2.3c, On the orthophoto plane, an image segmentation algorithm based on pixel color and texture features is used to divide the area into multiple regions and extract the region of interest. The brightness value histogram of all pixels in the region of interest is calculated, and the overall brightness and contrast are adjusted to a predefined standard level.

4. The method for evaluating tunnel blasting quality based on multi-source data fusion according to claim 1, characterized in that, In step S5, the machine learning model is a gradient boosting decision tree model; The training process gradually improves prediction accuracy by fitting the prediction residuals of previous models to each decision tree sequentially. The combination of hyperparameters of the model, including the maximum depth of the tree, the learning rate, and the number of subtrees, is determined by cross-validation combined with an automated parameter search strategy.

5. The tunnel blasting quality evaluation method based on multi-source data fusion according to claim 4, characterized in that, When training a gradient boosting decision tree model, the historical database is randomly divided into a subset for direct training, a subset for real-time performance verification and early stopping during training to prevent overfitting, and a test subset for evaluating the model's generalization ability.

6. The method for evaluating the quality of tunnel blasting based on multi-source data fusion according to claim 1, characterized in that, In step S4, after constructing the comprehensive feature vector and before inputting it into the model for training, each feature parameter in the vector is numerically scaled. The feature value is mapped to a range between zero and one by subtracting the minimum value of historical samples from the feature value of the comprehensive feature vector and then dividing by the range of historical samples.

7. The method for evaluating the quality of tunnel blasting based on multi-source data fusion according to claim 1, characterized in that, In step S5, the hyperparameters of the machine learning model are optimized and selected using the Bayesian optimization method. We construct a probabilistic surrogate model to fit the relationship between the hyperparameter combination and the model validation performance, and find the hyperparameter combination with the best validation performance through iteration.

8. The method for evaluating the quality of tunnel blasting based on multi-source data fusion according to claim 1, characterized in that, The visualization report described in step S6 includes side-by-side bar charts or radar charts; Each chart represents a key feature parameter, indicating the specific value of that parameter in this blasting and a reference range representing the numerical distribution of the parameter in the historical database of blasting samples of different grades.

9. The method for evaluating the quality of tunnel blasting based on multi-source data fusion according to claim 1, characterized in that, The method further includes a model update loop step S7: The system periodically checks newly accumulated blasting data samples whose quality levels have been confirmed by on-site engineers. The blasting data samples are added to the historical database, and the intelligent evaluation model for blasting quality is retrained or incrementally learned to update the model parameters to adapt to changes in geological conditions or blasting parameters during tunnel excavation.

Images