Intelligent identification and analysis method and system for shield tunnel face image in complex stratum

Abstract

This invention discloses a method and system for intelligent recognition and analysis of images from shield tunnel faces in complex geological formations, relating to the field of intelligent monitoring technology for shield tunnel construction. The method includes: a multi-source spoil image data acquisition step, acquiring endoscope images from the spoil chamber, video images from the spoil conveyor belt, and spoil sampling photos, and performing preprocessing; a spoil image multi-feature extraction and fusion step, extracting color, texture, and morphological features and fusing them through an attention mechanism; an intelligent formation type recognition step, outputting formation type, water content, and particle size distribution; an intelligent recommendation step for tunneling parameters, generating recommended values ​​for propulsion speed, cutterhead rotation speed, and spoil chamber pressure; and a recognition model feedback optimization step, updating model parameters based on actual tunneling results.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for shield tunnel construction, specifically to a method and system for intelligent recognition and analysis of shield tunnel face images in complex geological formations. Background Technology

[0002] As the mainstream construction method for urban underground space development, the shield tunneling method is widely used in projects such as subway tunnels, municipal utility tunnels, and cross-river tunnels. When the shield machine traverses complex geological formations, accurate identification of the geological conditions is crucial for the rational setting of tunneling parameters. When the shield machine traverses composite strata composed of various soil types such as clay, sand, gravel, and rock, the physical and mechanical properties of different strata vary significantly. Therefore, it is necessary to adjust tunneling parameters such as the advance speed, cutterhead rotation speed, and soil chamber pressure accordingly to ensure construction safety and efficiency.

[0003] Chinese patent CN120374708A discloses a machine vision-based method and device for calculating the volume of tunnel boring machine (TBM) excavated soil. This method uses a dual-modal input image segmentation network to extract and fuse features from real-time acquired target depth and RGB images. Based on the fused features, the target RGB image is segmented to obtain the coordinates of the excavated soil mask image and the excavated soil bounding box, thereby calculating the volume of the TBM excavated soil. This method primarily addresses the issue of accuracy in excavated soil volume calculation. Its technical solution focuses on dual-modal image segmentation and volume calculation, enabling real-time, high-precision measurement of TBM excavated soil volume.

[0004] However, the aforementioned existing technologies have the following technical problems: First, the method only focuses on the calculation of the volume of excavated soil and does not address the technical issue of identifying the current stratum type through excavated soil characteristics, thus failing to provide a stratum basis for adjusting tunneling parameters; second, the method only utilizes images of excavated soil truck compartments from a single source and does not integrate multi-source image data such as soil chamber endoscopes, excavated conveyor belts, and excavated soil sampling, resulting in a single source of image information and limited timeliness; third, the method does not establish a mapping relationship between excavated soil image features and stratum types, lacking the ability to comprehensively judge stratum types from multi-dimensional features such as excavated soil color, texture, and morphology; finally, the method does not provide a tunneling parameter recommendation function and cannot automatically generate appropriate tunneling parameter suggestions based on the identified stratum type.

[0005] Therefore, existing technologies cannot meet the needs of real-time identification of strata changes and intelligent recommendation of tunneling parameters in shield tunneling in complex strata. There is an urgent need for an intelligent method that can integrate multi-source slag and soil image data, identify strata types through multi-feature fusion, and provide recommendations for tunneling parameters. Summary of the Invention

[0006] To address the technical problems in existing technologies, such as lagging stratum identification, strong subjectivity in manual judgment of soil characteristics, and lack of intelligent recommendation of tunneling parameters, this invention provides a method and system for intelligent recognition and analysis of shield tunnel face images in complex strata.

[0007] The first aspect of this invention provides an intelligent recognition and analysis method for shield tunnel face images in complex strata, comprising: a multi-source spoil image data acquisition step, acquiring endoscope images of the soil chamber, video images of the spoil conveyor belt, and spoil sampling photos during the shield tunneling process, and obtaining geological survey data corresponding to the tunneling location; performing spatiotemporal alignment and normalization preprocessing on the acquired multi-source images to generate a standardized spoil image dataset; a spoil image multi-feature extraction and fusion step, extracting color feature vectors, texture feature vectors, and morphological feature vectors from each frame of the standardized spoil image dataset, and performing weighted fusion of the color feature vectors, texture feature vectors, and morphological feature vectors based on an adaptive attention mechanism to generate a fused feature vector; and a strata type step. The intelligent identification step inputs the fused feature vector into a pre-trained stratum type mapping model, outputting the stratum type classification result, water-bearing state identification result, and particle size distribution characteristics at the current tunneling location. The intelligent tunneling parameter recommendation step calculates and generates recommended values ​​for propulsion speed, cutterhead rotation speed, and soil chamber pressure based on the stratum type classification result, water-bearing state identification result, and particle size distribution characteristics through a stratum parameter association model. The identification model feedback optimization step obtains actual tunneling performance data after the tunnel boring machine performs tunneling operations according to the recommended propulsion speed, cutterhead rotation speed, and soil chamber pressure values, and updates the parameters of the stratum type mapping model and the stratum parameter association model based on the deviation between the actual tunneling performance data and the expected performance.

[0008] Preferably, the extraction of color feature vectors includes converting the image from the RGB color space to the HSV color space and the Lab color space respectively, and then performing joint feature extraction in both spaces.

[0009] Preferably, the adaptive attention mechanism calculates the similarity scores between multiple features and performs normalization to obtain attention weights, thereby achieving adaptive weighted fusion of features.

[0010] Preferably, the formation type mapping model adopts a multi-task learning architecture, and simultaneously outputs three types of identification results: formation type, water-bearing state, and particle size distribution.

[0011] Preferably, the online incremental learning algorithm uses a sliding time window and mini-batch gradient descent to continuously optimize the model parameters.

[0012] The second aspect of the present invention provides an intelligent recognition and analysis system for shield tunnel face images in complex strata, comprising: a multi-source spoil image data acquisition module, a spoil image multi-feature extraction and fusion module, a stratum type intelligent recognition module, a tunneling parameter intelligent recommendation module, and a recognition model feedback optimization module, with each module corresponding to a method step.

[0013] The beneficial effects of this invention include: by integrating multi-source image data such as soil chamber endoscope, muck conveyor belt, and muck sampling, the richness and reliability of stratum identification data are improved; by fusing multiple features such as color, texture, and morphology, the limitations of single feature identification are overcome, and the accuracy of stratum type identification reaches over 90%; by establishing a correlation model between stratum type and tunneling parameters, intelligent recommendation of tunneling parameters is realized, reducing the subjectivity of human experience judgment; and through an online incremental learning mechanism, the model can adapt to the stratum characteristics of different projects, improving the system's generalization ability. Attached Figure Description

[0014] Figure 1 This is a flowchart of the intelligent recognition and analysis method for shield tunnel face images in complex strata according to the present invention.

[0015] Figure 2 This is an architecture diagram of the intelligent recognition and analysis system for shield tunnel face images in complex geological formations, as described in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0017] See Figure 1 The intelligent recognition and analysis method for shield tunnel face images in complex strata provided by this invention includes five core steps: multi-source spoil image data acquisition step S1, multi-feature extraction and fusion of spoil images step S2, intelligent identification of strata type step S3, intelligent recommendation of tunneling parameters step S4, and feedback optimization of the recognition model step S5. 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 feedback result of step S5 inversely influences the parameters of the model in steps S2 and S3, achieving continuous optimization of recognition accuracy.

[0018] Step S1: Multi-source construction waste image data acquisition steps.

[0019] The multi-source excavation image data acquisition step is used to collect excavation image data from multiple sources during shield tunneling, and to perform spatiotemporal alignment and preprocessing.

[0020] In this embodiment, the image data sources include three categories: The first category is endoscope images from the soil chamber, acquired through an industrial endoscope camera installed inside the tunnel boring machine's soil chamber. This camera features a dustproof and waterproof design, a resolution of 1920×1080 pixels, and a frame rate of 25 frames per second, enabling real-time observation of the soil condition within the soil chamber. The second category is video images from the muck discharge conveyor, acquired through a high-definition camera mounted above the muck discharge conveyor. This camera has a resolution of 1280×720 pixels and a frame rate of 30 frames per second, allowing continuous monitoring of the muck flow discharged from the soil chamber. The third category is photographs of muck sampling, taken by on-site technicians according to a pre-set sampling procedure under standard lighting conditions. These photographs have a resolution of 2048×1536 pixels. In addition, pre-construction geological survey data is also required, including information on the depth distribution, thickness, and soil type of each stratum.

[0021] Spatiotemporal alignment of multi-source images is a crucial step in ensuring the effectiveness of data fusion. In this embodiment, a spatiotemporal indexing method based on tunneling mileage and acquisition timestamps is used for alignment. Specifically, a two-dimensional index table is established with tunneling mileage as the primary index and acquisition time as the secondary index. For the same tunneling location, there is a time delay between the discharge of excavated soil from the soil bin and its transport by the conveyor belt, which is related to the length and operating speed of the conveyor belt.

[0022] Assume the acquisition time of the endoscope image in the earth chamber is... The corresponding tunneling mileage is The belt conveyor length is The belt running speed is The corresponding time of the video image of the slag discharge conveyor belt. The calculation is as follows:

[0023] ,

[0024] in, The processing time for mixing and discharging excavated soil in the soil chamber is empirically calibrated based on the tunnel boring machine model and tunneling speed, typically ranging from 30 to 120 seconds. When an image source is missing at a corresponding time point, the nearest-time interpolation method is used to fill in the gaps, i.e., the adjacent frame image with the smallest time difference is selected as the replacement.

[0025] Image preprocessing comprises three sub-steps: image enhancement, size normalization, and color correction. Image enhancement employs a contrast-limited adaptive histogram equalization method, dividing the image into 8×8 sub-blocks and performing histogram equalization on each block separately. The contrast limit coefficient is set to 2.0 to enhance local image contrast and suppress noise amplification. Size normalization scales all images to a uniform size of 256×256 pixels, using bilinear interpolation to maintain image smoothness. Color correction addresses color deviations from different cameras by adjusting white balance using a pre-calibrated color correction matrix, ensuring color consistency across multiple image sources.

[0026] The preprocessed multi-source image data is organized in order of tunneling mileage to form a standardized muck image dataset. Each sample contains three types of images: endoscope images of the soil chamber, video images of the muck conveyor belt, and photos of muck sampling at the same tunneling location, as well as prior geological exploration information for the corresponding location.

[0027] Step S2: Multi-feature extraction and fusion of slag and soil images.

[0028] The multi-feature extraction and fusion step of the slag image is used to extract multi-dimensional features from the preprocessed slag image and perform feature fusion through an adaptive attention mechanism.

[0029] The color feature vector extraction employs a joint HSV-Lab color space method. First, the RGB image is converted to the HSV color space. The HSV color space decomposes color information into three channels: hue (H), saturation (S), and lightness (V), which better aligns with human color perception. Histogram statistics are performed on the H channel across 30 intervals, and on the S and V channels respectively across 16 intervals, resulting in a 62-dimensional HSV histogram feature. Simultaneously, the RGB image is converted to the Lab color space. In the Lab color space, the L channel represents lightness, the a channel represents the red-green color pair, and the b channel represents the yellow-blue color pair. The mean and standard deviation of the a and b channels are calculated, forming a 4-dimensional Lab statistical feature. The HSV histogram feature is concatenated with the Lab statistical feature, and then principal component analysis is used to reduce the dimensionality to 128 dimensions, forming the final color feature vector. .

[0030] Different types of strata exhibit different color characteristics in their slag: clay slag is usually yellowish-brown or grayish-brown, with positive values ​​in both the a and b channels in the Lab color space; sandy slag is lighter in color, appearing grayish-yellow or grayish-white, with low saturation; gravel slag has mixed colors, containing multiple color components, and its color histogram is relatively evenly distributed; the color of rock strata slag is related to the rock type, such as granite slag being grayish-white with black spots, and sandstone slag being reddish-brown or yellowish-brown.

[0031] Texture feature vector extraction employs the gray-level co-occurrence matrix (GLCM) method. First, the color image is converted to grayscale, and the grayscale values ​​are quantized to 64 levels to reduce computational cost. The GLCM is calculated at four directions (0°, 45°, 90°, 135°) and four distances (1 pixel, 2 pixels, 4 pixels, 8 pixels). Four statistics are extracted from each GLCM: contrast, correlation, energy, and homogeneity.

[0032] Contrast reflects the magnitude of the gray-level difference between adjacent pixels in an image, and is calculated using the following formula:

[0033] ,

[0034] in, This represents the number of grayscale levels, with a value of 64. Position in the gray-level co-occurrence matrix The normalized probability value at the specified location represents the grayscale value. The pixel value is determined by the distance and orientation parameters of the neighboring pixels. The joint probability.

[0035] Correlation reflects the degree of linear correlation between gray values ​​in an image in a certain direction. The calculation formula is:

[0036] ,

[0037] in, and These are the mean values ​​of the gray-level co-occurrence matrix along the row and column directions, respectively. and This represents the corresponding standard deviation.

[0038] Energy reflects the uniformity of image texture, and the calculation formula is:

[0039] ,

[0040] Homogeneity reflects the local uniformity of image texture, and is calculated using the following formula:

[0041] ,

[0042] The combination of four directions and four distances generates 16 gray-level co-occurrence matrices. Four statistics are extracted from each matrix, resulting in a total of 64-dimensional features. These 64-dimensional features are then standardized and mapped to 256 dimensions through a fully connected layer to form a texture feature vector. .

[0043] Different types of strata exhibit different texture characteristics in their slag: clay slag has a fine texture, relatively uniform texture, low contrast, and high homogeneity; sandy slag has distinct particles, relatively coarse texture, and medium contrast; pebble slag contains large particles, has uneven texture, high contrast, and low energy; the texture of rock slag is related to the degree of rock weathering, with strongly weathered rock slag having a texture similar to sand, while fresh rock slag shows obvious boundaries of rock fragments.

[0044] The extraction of morphological feature vectors is based on the contour analysis of slag particles. First, edge detection is performed on the preprocessed image using the Canny edge detection algorithm, with a high threshold set to 200 and a low threshold set to 100. Contour extraction is then performed on the detected edges, filtering out noisy contours with an area less than 100 pixels. For each valid contour, four morphological parameters are calculated: area A, perimeter P, roundness C, and aspect ratio R.

[0045] Circularity describes how closely a contour shape approximates a circle, and the formula for calculation is:

[0046] ,

[0047] in, This represents the outline area, expressed in squared pixels. This represents the perimeter of the outline, in pixels. The roundness value ranges from 0 to 1, with values ​​closer to 1 indicating a shape that is closer to a circle.

[0048] The aspect ratio is obtained by calculating the minimum bounding rectangle of the outline. The calculation formula is:

[0049] ,

[0050] in, and These are the width and height of the smallest bounding rectangle, respectively.

[0051] The area, perimeter, roundness, and aspect ratio distributions of all slag particles in the statistical image were analyzed. Each parameter was statistically analyzed using histograms across 16 intervals, resulting in a 64-dimensional feature vector. .

[0052] Multi-feature fusion employs an adaptive attention mechanism, which dynamically adjusts the contribution weights of various features based on different geological conditions. The three types of feature vectors are transformed linearly to generate a query vector Q, a key vector K, and a value vector V.

[0053] ,

[0054] ,

[0055] ,

[0056] in, , , Let be learnable linear transformation weight matrices, where the superscripts c, t, and m represent color, texture, and morphological features, respectively. All transformation matrices map their respective feature vectors to a unified dimension. .

[0057] Calculate the attention score matrix between features:

[0058] ,

[0059] in, This represents a vector concatenation operation. For the attention score matrix, each element Indicates the first Class features for the first The level of attention given to class features.

[0060] The attention weight matrix is ​​obtained by performing Softmax normalization on the attention score matrix:

[0061] ,

[0062] The Softmax operation normalizes the data row by row, making the sum of the elements in each row equal to 1.

[0063] Weighted fusion of value vectors based on attention weight matrix:

[0064] ,

[0065] in, To fuse feature vectors, information from three types of features—color, texture, and shape—is integrated, and the contribution weight of each feature is adaptively determined by an attention mechanism.

[0066] During the training phase, the parameters of the attention mechanism are jointly optimized with the subsequent classification network. During the inference phase, the attention weights reflect the importance of different features to the current stratum identification. For example, when the moisture content of the slag is high, making color features more prominent, the attention mechanism will automatically increase the weight of color features; when the slag particles are large, making morphological features more distinguishable, the weight of morphological features will be increased accordingly.

[0067] Step S3: Intelligent identification of stratigraphic type.

[0068] The intelligent formation type identification step is used to identify the formation type, water-bearing state, and particle size distribution characteristics of the current tunneling location based on the fused feature vector.

[0069] The stratigraphic type mapping model employs a multi-task learning architecture, including a shared feature encoder and three task-specific classification or regression heads. The feature encoder uses a residual convolutional network structure, containing four residual blocks. Each residual block consists of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, and includes skip connections to mitigate the vanishing gradient problem. The feature encoder further encodes the 512-dimensional fused feature vector into 256-dimensional deep semantic features. .

[0070] The stratigraphic type classification branch is used to identify the stratigraphic type corresponding to the current waste soil. Considering that multiple soil components may coexist in complex strata, a multi-label classification method is adopted. The classification branch contains two fully connected layers. The first layer maps 256-dimensional features to 128 dimensions, and the second layer maps 128-dimensional features to a 4-dimensional output, corresponding to four stratigraphic types: clay, sand, gravel, and rock. The output layer uses the Sigmoid activation function, and each output value represents the probability of the corresponding stratigraphic type existing.

[0071] ,

[0072] in, For the Sigmoid function, , This is the weight matrix. , For bias vectors, This is the probability vector for stratigraphic types.

[0073] When the probability value is greater than the preset classification confidence threshold (0.5), the stratigraphic type is determined to exist. The proportion of each stratigraphic type is calculated through probability normalization:

[0074] ,

[0075] in, For the first The proportion of different soil strata types This represents the corresponding probability value.

[0076] The moisture content classification branch is used to identify the moisture content level of the waste soil. Moisture content has a significant impact on the color and flowability of the waste soil, and the moisture content is divided into four levels: dry, slightly wet, moist, and saturated. The classification branch structure is similar to that of the stratigraphic type classification branch, and the output layer uses the Softmax activation function for single-label classification.

[0077] ,

[0078] in, Given a probability vector for water content, the category with the highest probability is selected as the water content identification result.

[0079] The particle size distribution regression branch is used to predict the particle size distribution characteristics of slag. Particle size distribution is represented by characteristic particle sizes, including... , , , Four feature sizes are provided, representing the particle size values ​​corresponding to cumulative passing mass percentages of 10%, 30%, 50%, and 60%, respectively. The regression branch contains two fully connected layers, outputting predicted values ​​for the four feature sizes:

[0080] ,

[0081] in, This is the feature particle size prediction vector, in millimeters.

[0082] The non-uniformity coefficient can be further calculated based on the characteristic particle size. and curvature coefficient :

[0083] ,

[0084] ,

[0085] The uniformity coefficient reflects the distribution range of soil particle size, while the curvature coefficient reflects the smoothness of the particle size distribution curve. According to soil mechanics principles, when... and At that time, the soil gradation was good.

[0086] The stratigraphic type mapping model is trained using a multi-task joint loss function:

[0087] ,

[0088] in, Binary cross-entropy loss for stratigraphic type classification, Cross-entropy loss for classifying water content states. This represents the mean squared error loss for particle size distribution regression. , , These are the weighting coefficients for each task, with values ​​of 1.0, 0.5, and 0.3 respectively.

[0089] The formula for calculating the stratigraphic type classification loss is:

[0090] ,

[0091] in, For batch size, For the first In the nth sample The true label (0 or 1) for each stratum type. This represents the corresponding predicted probability.

[0092] The model training data comes from images of excavated soil and geological survey records of historical tunnel boring machines. The training dataset contains 15,000 samples, covering various geological strata combinations. The Adam optimizer is used, with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. A cosine annealing learning rate scheduling strategy is employed to gradually decrease the learning rate during training. Data augmentation methods, including random horizontal flipping, random rotation, and random brightness adjustment, are used during training to improve the model's generalization ability.

[0093] In the test of this embodiment, the stratigraphic type mapping model achieved a stratigraphic type identification accuracy of 92.3% and a water-bearing state identification accuracy of 89.7% on the test set, with an average relative error of 8.5% for particle size distribution prediction.

[0094] Step S4: Intelligent recommendation of tunneling parameters.

[0095] The intelligent recommendation step for tunneling parameters is used to generate suitable recommended values ​​for advance speed, cutterhead rotation speed, and soil chamber pressure based on the identified geological features.

[0096] Recommended tunneling parameters include three key parameters: propulsion speed. Cutter head speed and earth pressure The optimal tunneling parameters vary significantly under different geological conditions: clay soil has good plasticity and can use a higher advance speed; sandy soil has poor stability and requires an appropriate reduction in advance speed while maintaining stable soil pressure; gravel soil causes greater wear on the cutterhead and requires a reduction in cutterhead speed and control of the advance force; rock soil has high tunneling resistance and requires an increase in cutterhead torque and control of the advance speed.

[0097] The stratigraphic parameter association model uses the proportion vector of each stratigraphic type in the stratigraphic type classification results. Moisture content level corresponding to the water content status identification result (Corresponding to dry, slightly moist, moist, and saturated conditions respectively) and characteristic particle radial quantity in particle size distribution characteristics. As input features, the recommended values ​​for propulsion speed, cutterhead rotation speed, and soil chamber pressure are output through a multilayer perceptron network.

[0098] The input feature vector is constructed as follows: the stratigraphic type proportion vector is directly represented using 4D; the water content grade is converted into a 4D vector using one-heat encoding; the radial quantity of the feature particles is logarithmically transformed and then standardized to form a 4D vector. These features are then concatenated to form a 12-dimensional input feature vector. .

[0099] The multilayer perceptron network contains three hidden layers with 64, 32, and 16 neurons respectively, all using ReLU activation. The output layer contains three neurons, outputting recommended values ​​for propulsion speed, cutterhead rotation speed, and soil chamber pressure. To ensure the output values ​​remain within reasonable ranges, upper and lower limits are set for each parameter:

[0100] ,

[0101] ,

[0102] ,

[0103] in, , , This is the original output from the network. The Sigmoid function maps its output to the (0,1) interval. , The lower and upper limits of the propulsion speed are set at 10 mm / min and 80 mm / min, respectively. , These are the lower and upper limits of the cutter head rotation speed, with values ​​of 0.5 rpm and 2.5 rpm respectively. , These are the lower and upper limits of the earth pressure chamber, with values ​​of 0.8 times the at-rest earth pressure and 1.3 times the at-rest earth pressure, respectively.

[0104] The training data for the geological parameter correlation model comes from historical tunnel boring machine (TBM) records. Data from tunneling loops with good performance (i.e., surface settlement less than the warning value, smooth progress without abnormal shutdowns) are selected as training samples. The corresponding geological features are used as input, and the actual tunneling parameters used are used as labels. The training loss uses mean squared error.

[0105] ,

[0106] in, For the sample size, , , For the first The actual tunneling parameters of each sample , , This is the corresponding predicted value.

[0107] In practical applications, the recommended values ​​for tunneling parameters serve as reference suggestions for operators, who can make appropriate adjustments based on the actual site conditions. The system also displays the confidence interval of the recommended values; when the confidence level of the identification results is low, the suggested range of the recommended values ​​is expanded.

[0108] Step S5: Identify model feedback optimization steps

[0109] The identification model feedback optimization step is used to optimize the parameters of the formation type mapping model and the formation parameter association model based on the actual tunneling results, forming a closed-loop control mechanism.

[0110] Actual tunneling performance data includes three categories of indicators: tunneling efficiency indicators, construction safety indicators, and equipment status indicators. Tunneling efficiency indicators include the deviation between the actual advance speed and the recommended value, and the tunneling time per ring; construction safety indicators include surface settlement monitoring values ​​and the range of soil chamber pressure fluctuations; equipment status indicators include cutterhead torque, propulsion cylinder pressure, and screw conveyor speed.

[0111] The expected results are set based on engineering design requirements and safety specifications: surface settlement is controlled within the warning value (usually 20 mm), soil chamber pressure fluctuation is controlled within ±10% of the set value, and cutterhead torque is maintained within 60% to 80% of the rated value.

[0112] The deviation value is calculated using a comprehensive evaluation method. Let the actual tunneling effect data vector be... The expected effect vector is The deviation value is calculated as follows:

[0113] ,

[0114] in, For the number of performance indicators, These are the weighting coefficients for each indicator. and The first The actual and expected values ​​of each indicator. The weights of each indicator are set according to the importance of the project: the construction safety indicator has a weight of 0.5, the tunneling efficiency indicator has a weight of 0.3, and the equipment status indicator has a weight of 0.2.

[0115] Online incremental learning algorithms use a sliding time window mechanism to collect the most recent data. A sample of tunneling operations in rings (valued at 50 rings). The deviation value of a given sample... Exceeding the preset error threshold When the value is 0.15, the sample is added to the update queue. When the number of samples in the update queue reaches the preset batch size (value 16), the model parameters are updated.

[0116] The model parameters are updated using mini-batch gradient descent. For the stratigraphic type mapping model, samples in the update queue are used as input, and manually corrected stratigraphic labels or inversely derived stratigraphic labels are used as supervision signals to calculate the classification loss and update the network parameters. The method for inversely derived stratigraphic labels is to infer the most likely combination of stratigraphic types based on actual tunneling performance data and known stratigraphic-parameter relationships.

[0117] The learning rate for parameter updates is set to a small value (0.0001) to avoid drastic changes in model parameters that could lead to unstable recognition performance. Simultaneously, an elastic weight consolidation method is employed to prevent catastrophic forgetting, applying regularization constraints to parameters historically important during parameter updates.

[0118] ,

[0119] in, For the loss of new samples, The current parameter value. The parameter values ​​before the update. These are the diagonal elements of the Fisher information matrix, measuring the importance of this parameter to historical tasks. This is the regularization coefficient, with a value of 1000.

[0120] By identifying model feedback optimization steps, the stratigraphic type mapping model and the stratigraphic parameter association model can adaptively adjust according to the actual situation of the current project, improving the accuracy of identification and recommendation. In the long-term operation test of this embodiment, after online learning of 500 rings of tunneling data, the accuracy of stratigraphic type identification increased from the initial 92.3% to 95.1%, and the average deviation of parameter recommendation decreased by 23%.

[0121] See Figure 2 The intelligent recognition and analysis system for shield tunnel face images in complex strata provided by the present invention includes a multi-source slag image data acquisition module 1, a slag image multi-feature extraction and fusion module 2, a stratum type intelligent recognition module 3, a tunneling parameter intelligent recommendation module 4, and a recognition model feedback optimization module 5.

[0122] The multi-source muck image data acquisition module 1 is used to acquire endoscope images of the muck chamber, video images of the muck conveyor belt, and photos of muck sampling during the shield tunneling process. It also acquires geological survey data corresponding to the tunneling location, performs spatiotemporal alignment and normalization preprocessing on the acquired multi-source images, and generates a standardized muck image dataset. The multi-source muck image data acquisition module 1 includes an image acquisition unit, a data synchronization unit, and a preprocessing unit. The image acquisition unit communicates with the endoscope camera in the muck chamber, the conveyor belt camera, and the sample photo upload terminal to receive multi-source image data in real time. The data synchronization unit establishes a spatiotemporal index for the multi-source images based on the tunneling mileage and acquisition timestamps to achieve image data alignment. The preprocessing unit enhances, normalizes, and corrects the color of the aligned images. The specific working principle and parameter settings of the multi-source muck image data acquisition module 1 are the same as those described in step S1 of the method embodiment.

[0123] The multi-feature extraction and fusion module 2 for construction waste images extracts color feature vectors, texture feature vectors, and morphological feature vectors from each frame of the standardized construction waste image dataset. Based on an adaptive attention mechanism, it weights and fuses these feature vectors to generate a fused feature vector. The module 2 includes a color feature extraction unit, a texture feature extraction unit, a morphological feature extraction unit, and an attention fusion unit. The color feature extraction unit uses the HSV-Lab dual color space joint method to extract color feature vectors. The texture feature extraction unit uses the gray-level co-occurrence matrix method to extract texture feature vectors. The morphological feature extraction unit extracts morphological feature vectors based on edge detection and contour analysis. The attention fusion unit achieves weighted fusion of multiple features through an adaptive attention mechanism. The specific working principle, feature extraction algorithm, and fusion formula of the multi-feature extraction and fusion module 2 are the same as those described in step S2 of the method embodiment.

[0124] The intelligent stratigraphic type identification module 3 is used to input the fused feature vector into a pre-trained stratigraphic type mapping model and output the stratigraphic type classification result, water-bearing state identification result, and particle size distribution feature at the current tunneling location. The intelligent stratigraphic type identification module 3 includes a feature encoder, a stratigraphic type classification unit, a water-bearing state classification unit, and a particle size distribution regression unit. The feature encoder uses a residual convolutional network structure to perform deep semantic encoding on the fused feature vector. The stratigraphic type classification unit uses a multi-label classification method to output the proportions of four stratigraphic types: clay, sand, gravel, and rock. The water-bearing state classification unit uses a single-label classification method to output four water-bearing levels: dry, slightly wet, wet, and saturated. The particle size distribution regression unit outputs the predicted feature particle size. The specific working principle, network structure, and loss function of the intelligent stratigraphic type identification module 3 are the same as those described in step S3 of the method embodiment.

[0125] The intelligent recommendation module 4 for tunneling parameters calculates and generates recommended values ​​for advance speed, cutterhead rotation speed, and soil chamber pressure based on the stratigraphic type classification results, water-bearing state identification results, and particle size distribution characteristics using a stratigraphic parameter association model. The intelligent recommendation module 4 includes a feature construction unit and a parameter recommendation unit. The feature construction unit concatenates the stratigraphic type proportion vector, the unique thermal encoding of the water-bearing level, and the characteristic particle radial quantity to form an input feature vector. The parameter recommendation unit uses a multilayer perceptron network to output recommended values ​​for advance speed, cutterhead rotation speed, and soil chamber pressure. The specific working principle, network structure, and constraint range of the intelligent recommendation module 4 for tunneling parameters are the same as those described in step S4 of the method embodiment.

[0126] The identification model feedback optimization module 5 is used to acquire actual tunneling performance data after the tunnel boring machine (TBM) performs tunneling operations according to recommended values ​​for propulsion speed, cutterhead rotation speed, and soil chamber pressure. It updates the parameters of the stratum type mapping model and the stratum parameter association model based on the deviation between the actual and expected performance data. The identification model feedback optimization module 5 includes an performance acquisition unit, a deviation calculation unit, and a parameter update unit. The performance acquisition unit communicates with the TBM's sensing system to collect real-time tunneling efficiency indicators, construction safety indicators, and equipment status indicators. The deviation calculation unit calculates the deviation between the actual and expected performance using a comprehensive evaluation method. The parameter update unit updates the model parameters using an online incremental learning algorithm and an elastic weight consolidation method. The specific working principle, deviation calculation formula, and parameter update algorithm of the identification model feedback optimization module 5 are the same as those described in step S5 of the method embodiment.

[0127] The system of this invention comprises two parts: edge computing nodes and a cloud server. The edge computing nodes are deployed on the tunnel boring machine (TBM) and are responsible for real-time tasks such as image acquisition, preprocessing, and feature extraction. They are equipped with GPU acceleration modules to meet the computing power requirements of deep learning inference. The cloud server is responsible for model training, parameter updates, and historical data storage, and communicates with the edge computing nodes via industrial Ethernet. When the network connection is interrupted, the edge computing nodes can independently complete the functions of geological stratum identification and parameter recommendation; once the network is restored, the model parameters are automatically updated synchronously.

[0128] In the actual engineering application test of this invention, a tunnel section of a subway in a certain city was used as the test object. The total length of the section was 1200 meters, and the strata it traversed included a composite stratum of clay, sand, and gravel. During the test, a total of 800 rings of tunneling data were collected. The system's accuracy rate in identifying stratum type reached 93.5%, the accuracy rate in identifying water-bearing state reached 91.2%, and the adoption rate of recommended tunneling parameters reached 87.6%. Compared with the traditional manual experience-based judgment method, the tunneling efficiency assisted by the system was improved by 15.3%, and the surface settlement control effect was improved by 18.7%.

[0129] 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 intelligent recognition and analysis of images of shield tunnel faces in complex geological formations, characterized in that, include: The multi-source slag image data acquisition steps include acquiring endoscope images of the soil chamber, video images of the slag conveyor belt, and photos of slag sampling during the shield tunneling process, and obtaining geological survey data for the corresponding tunneling locations. The acquired multi-source images are then preprocessed with spatiotemporal alignment and normalization to generate a standardized slag image dataset. The multi-feature extraction and fusion step of the construction waste image involves extracting color feature vectors, texture feature vectors, and morphological feature vectors from each frame of the standardized construction waste image dataset, and then weighting and fusing the color feature vectors, texture feature vectors, and morphological feature vectors based on an adaptive attention mechanism to generate a fused feature vector. The intelligent stratigraphic type identification step involves inputting the fused feature vector into a pre-trained stratigraphic type mapping model, and outputting the stratigraphic type classification result, water-bearing state identification result, and particle size distribution characteristics at the current tunneling location. The stratigraphic type classification result includes one or more combinations of clay, sand, gravel, and rock strata and their corresponding proportions. The intelligent recommendation step for tunneling parameters involves calculating and generating recommended values ​​for advance speed, cutterhead rotation speed, and soil chamber pressure based on the geological type classification results, the water-bearing state identification results, and the particle size distribution characteristics through a geological parameter association model. The identification model feedback optimization steps involve obtaining actual tunneling performance data after the tunnel boring machine performs tunneling operations according to the recommended values ​​of the propulsion speed, cutterhead rotation speed, and soil chamber pressure; calculating the deviation between the actual tunneling performance data and the expected performance; and updating the parameters of the stratum type mapping model and the stratum parameter association model based on the deviation value using an online incremental learning algorithm.

2. The intelligent recognition and analysis method for shield tunnel face images in complex strata according to claim 1, characterized in that, The extraction of the color feature vector includes: converting the preprocessed image from the RGB color space to the HSV color space and the Lab color space respectively; calculating the histogram distribution of the H channel, S channel and V channel in the HSV color space; calculating the mean and standard deviation of the a channel and b channel in the Lab color space; and concatenating the histogram distribution, the mean and the standard deviation to form the color feature vector.

3. The intelligent recognition and analysis method for shield tunnel face images in complex strata according to claim 1, characterized in that, The weighted fusion of the color feature vector, texture feature vector, and morphological feature vector based on the adaptive attention mechanism includes: inputting the color feature vector, texture feature vector, and morphological feature vector into the corresponding fully connected layers to generate a query vector, a key vector, and a value vector, respectively; calculating the similarity score between the query vector and the key vector; normalizing the similarity score to obtain attention weights; and weighting and summing the value vectors based on the attention weights to obtain the fused feature vector.

4. The intelligent recognition and analysis method for shield tunnel face images in complex strata according to claim 1, characterized in that, The stratigraphic type mapping model includes a feature encoder and a multi-task classification head. The feature encoder uses a residual convolutional network structure to extract deep semantic features. The multi-task classification head includes a stratigraphic type classification branch, a water-bearing state classification branch, and a particle size distribution regression branch. The three branches share the output features of the feature encoder.

5. The intelligent recognition and analysis method for shield tunnel face images in complex strata according to claim 1, characterized in that, The formation parameter association model uses the proportion of each formation type in the formation type classification result, the water content level corresponding to the water content status identification result, and the characteristic particle size in the particle size distribution feature as input features, and outputs the recommended values ​​of propulsion speed, cutterhead rotation speed, and soil chamber pressure through a multilayer perceptron network.

6. The intelligent recognition and analysis method for shield tunnel face images in complex strata according to claim 1, characterized in that, The extraction of the texture feature vector includes: performing grayscale processing on the preprocessed image, calculating the grayscale co-occurrence matrix under multiple preset directions and multiple distance parameters, extracting four texture statistics—contrast, correlation, energy, and homogeneity—from the grayscale co-occurrence matrix, and concatenating the texture statistics under different directions and distances to form the texture feature vector.

7. The intelligent recognition and analysis method for images of shield tunnel faces in complex strata according to claim 1, characterized in that, The extraction of the morphological feature vector includes: performing edge detection on the preprocessed image to obtain the outline of the slag particles; calculating the area, perimeter, roundness, and aspect ratio of each slag particle outline; statistically analyzing the distribution histogram of the area, perimeter, roundness, and aspect ratio of all slag particles; and flattening the distribution histogram to form the morphological feature vector.

8. The intelligent recognition and analysis method for images of shield tunnel faces in complex geological formations according to claim 1, characterized in that, The online incremental learning algorithm includes: setting a sliding time window to collect a preset number of recent tunneling operation samples, calculating the recognition error of each sample, triggering model parameter updates when the recognition error exceeds a preset error threshold, and using mini-batch gradient descent to update the network weights in the formation type mapping model and the formation parameter association model.

9. The intelligent recognition and analysis method for images of shield tunnel faces in complex strata according to claim 1, characterized in that, The spatiotemporal alignment of the multi-source images includes: establishing a spatiotemporal index of the multi-source images based on the tunneling mileage of the tunnel boring machine and the acquisition timestamp; registering the endoscope images of the soil chamber, the video images of the muck conveyor belt, and the muck sampling photos at the same tunneling location; and interpolating to complete the image source when a certain image source is missing.

10. A system for intelligent recognition and analysis of images of shield tunnel faces in complex geological formations, used to implement the method for intelligent recognition and analysis of images of shield tunnel faces in complex geological formations as described in any one of claims 1-9, characterized in that, include: The multi-source slag image data acquisition module is used to acquire endoscope images of the soil chamber, video images of the slag conveyor belt, and photos of slag sampling during the shield tunneling process, obtain geological survey data of the corresponding tunneling location, perform spatiotemporal alignment and normalization preprocessing on the acquired multi-source images, and generate a standardized slag image dataset. The multi-feature extraction and fusion module for construction waste images is used to extract color feature vectors, texture feature vectors, and morphological feature vectors from each frame of the standardized construction waste image dataset, and to perform weighted fusion of the color feature vectors, texture feature vectors, and morphological feature vectors based on an adaptive attention mechanism to generate a fused feature vector. The intelligent stratigraphic type identification module is used to input the fused feature vector into a pre-trained stratigraphic type mapping model and output the stratigraphic type classification result, water-bearing state identification result, and particle size distribution characteristics at the current tunneling location. The intelligent recommendation module for tunneling parameters is used to calculate and generate recommended values ​​for advance speed, cutterhead rotation speed and soil chamber pressure based on the geological type classification results, the water-bearing state identification results and the particle size distribution characteristics through the geological parameter association model. The identification model feedback optimization module is used to obtain the actual tunneling effect data after the tunnel boring machine performs tunneling operations according to the recommended values ​​of the propulsion speed, the cutterhead rotation speed, and the soil chamber pressure, and to update the parameters of the stratum type mapping model and the stratum parameter association model according to the deviation between the actual tunneling effect data and the expected effect.

Images