Method and device for evaluating the compressive strength of pipes by hyperspectral and rebound value correlation calibration

Abstract

This invention relates to the field of urban underground infrastructure inspection and structural safety assessment technology, and discloses a method and equipment for assessing the compressive strength of pipelines by hyperspectral and rebound value correlation calibration. The method includes: setting up rebound value measuring points in the accessible area of ​​the drainage pipeline, collecting rebound values ​​and hyperspectral images containing the measuring points; preprocessing the hyperspectral images to construct a calibration sample dataset correlating the hyperspectral images and rebound values; using this dataset to train a rebound value mapping model based on hyperspectral features to achieve the conversion from hyperspectral images to rebound values; collecting hyperspectral images of the inner wall surface of the pipeline and inputting them into the trained rebound value mapping model to generate rebound values ​​at the corresponding locations; combining the conversion relationship between rebound values ​​and compressive strength to calculate the concrete compressive strength, and generating the spatial distribution result of the pipeline inner wall compressive strength. This invention achieves high-precision rebound value prediction and automated, efficient assessment of pipeline compressive strength through hyperspectral and rebound value correlation calibration.

Description

Technical Field

[0001] This invention relates to the field of urban underground infrastructure inspection and structural safety assessment technology, specifically to a method and equipment for assessing the compressive strength of pipelines based on hyperspectral and rebound value correlation calibration. Background Technology

[0002] With the continuous advancement of urbanization, the scale of urban drainage pipeline networks continues to expand. Concrete drainage pipelines with long service lives commonly suffer from aging, corrosion, freeze-thaw damage, chemical erosion, and fatigue damage, leading to a decline in the structural load-bearing capacity of the pipelines, frequent local collapses and leaks, and posing serious threats to urban operational safety. Therefore, regular testing and evaluation of key mechanical properties of drainage pipeline concrete, such as compressive strength, is of significant engineering importance. Currently, concrete mechanical property testing in engineering practice mainly includes core sampling and rebound testing. While core sampling offers high accuracy, it is a destructive testing method, which suffers from drawbacks such as damaging the structural integrity at sampling points, high construction costs, complex operation, and unsuitability for confined spaces, making it unsuitable for continuous testing of large-scale pipeline structures. Rebound testing, as a commonly used non-destructive testing method, offers advantages such as ease of operation and high efficiency. However, this method relies on manual on-site operation and is only suitable for areas around manholes or directly accessible locations. It struggles to effectively test deep within pipelines, in curved sections, and in confined spaces, thus limiting its testing scope. Summary of the Invention

[0003] To address the aforementioned shortcomings in the existing technology, this invention provides a method and equipment for evaluating the compressive strength of pipelines by correlating hyperspectral data with rebound values, thereby solving the problems of low efficiency and accuracy in pipeline compressive strength testing in existing methods.

[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for assessing the compressive strength of pipes based on hyperspectral correlation and rebound value calibration includes the following steps: Rebound value measuring points were set up in the accessible area of ​​the drainage pipe to collect rebound value data, and hyperspectral images containing the rebound value measuring points were also collected. After preprocessing the hyperspectral images, a calibration sample dataset is constructed by combining the rebound value data. Using a calibrated sample dataset, a rebound value mapping model based on hyperspectral images is trained to generate a trained rebound value mapping model based on hyperspectral images, thereby realizing the conversion from hyperspectral images to rebound values. Hyperspectral images of the inner wall surface of the drainage pipe to be tested are acquired. After data preprocessing, the images are input into a trained rebound value mapping model based on hyperspectral images to generate rebound values ​​at corresponding locations in the inner wall region of the drainage pipe. Based on the rebound values ​​at corresponding locations on the inner wall of the drainage pipe, and combined with the conversion relationship between rebound values ​​and compressive strength, the compressive strength of concrete is calculated, generating the spatial distribution results of the compressive strength of the inner wall of the drainage pipe.

[0005] A device for acquiring hyperspectral images includes a hyperspectral camera 1, a front-end searchlight and a projection camera group 2, and an elastic wave pipeline inspection robot.

[0006] The present invention has the following beneficial effects: The proposed method for assessing the compressive strength of pipes by correlating hyperspectral images with rebound values ​​involves acquiring hyperspectral images and rebound values, calibrating the rebound values ​​with the hyperspectral images, generating calibration sample pairs, and training a rebound value mapping model. This allows for high-precision prediction of rebound values ​​at corresponding locations on the inner wall of the drainage pipe after obtaining a hyperspectral image of the pipe's inner surface. Simultaneously, based on the predicted rebound values ​​at these locations, a conversion between rebound values ​​and compressive strength is performed. By calculating the concrete compressive strength, a stable and reliable spatial distribution result of the compressive strength of the drainage pipe's inner wall is generated, improving the accuracy and efficiency of the drainage pipe's inner wall compressive strength detection. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the pipeline compressive strength assessment method based on the correlation between hyperspectral data and rebound value proposed in this invention. Figure 2 This is a schematic diagram of the rebound value mapping model based on hyperspectral features in the embodiment; Figure 3 This is a schematic diagram of the structure of a device proposed in this invention. Detailed Implementation

[0008] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0009] The specific embodiments of this invention are as follows: like Figure 1 As shown, the method for assessing the compressive strength of pipes by correlating hyperspectral data with rebound values ​​includes the following steps: Step 1: Set up rebound value measurement points in the accessible area of ​​the drainage pipe, collect rebound value data, and simultaneously acquire hyperspectral images containing the rebound value measurement points.

[0010] This step involves acquiring rebound value data and hyperspectral images containing rebound value measurement points, so that subsequent steps can construct a calibration sample dataset with a one-to-one correspondence between hyperspectral images and rebound values; the operation is as follows: First, in the accessible area of ​​the drainage pipes, i.e., near the pipe inspection wells, where rebound testing can be performed directly, deploy as many rebound value measurement points as possible. To avoid the distribution of a single dataset being offset, rebound value data should be collected at least 300 test points at the wellheads of three drainage pipe inspection wells every 100 meters.

[0011] Secondly, a rebound hammer was used to measure the rebound value and collect several rebound value data. At the same time, with the rebound value test point as the core, a hyperspectral camera was used to capture a hyperspectral image containing the rebound value test point under the illumination of a hyperspectral light source. When collecting the hyperspectral image containing the rebound value test point, in order to avoid measurement errors caused by insufficient illumination in the edge area, the rebound value test point should be far away from the edge of the image.

[0012] Before acquiring rebound data and hyperspectral images, the rebound hammer, hyperspectral lamp source, and hyperspectral camera need to be calibrated: The rebound hammer should be calibrated using a standard anvil by striking it four times consecutively at the center of the anvil with the hammer facing downwards. The average value of the four strikes should be within 80 ± 2 mm to confirm that the rebound hammer is functioning correctly. The hyperspectral lamp source should be adjusted to a working wavelength of 2500 nm, with a recommended power of 100 W, to obtain a uniform illumination area. The hyperspectral camera's recognition range should be in the near-infrared band of 900 nm to 2500 nm. Before using the hyperspectral camera, a black-and-white calibration should be performed using the following formula:

[0013] In the formula, For pixels In the band with a wavelength of The reflectivity value at that location, , pixels x Axis coordinates y Axis coordinates; The pixels of the original hyperspectral image At wavelength The grayscale value at that location; For wavelength The changing white reference hyperspectral response value; For wavelength The changing hyperspectral response of the dark reference.

[0014] Furthermore, since the rebound value mapping model based on hyperspectral features proposed in this invention requires extensive training with a large amount of data in the early stages to achieve accurate matching between hyperspectral data and rebound values, this step involves acquiring a large training set of data from locations such as near the inspection well where it is easy to obtain a large number of hyperspectral images and rebound values ​​to complete model training. Based on this, in subsequent steps, the equipment proposed in this invention is used to collect hyperspectral image data for non-destructive testing deep within the pipeline. Specifically, since hyperspectral images are easily affected by different concrete mix proportions and different testing environments, training weights specific to that specific testing task should be trained first, and then the corresponding training weights should be used for testing. Therefore, this invention integrates a hyperspectral camera and a light source and mounts them on an existing pipeline robot, allowing the implementation of the method of this invention to be added to the existing pipeline robot.

[0015] Step 2: After preprocessing the hyperspectral images, combine them with the rebound value data to construct a calibration sample dataset.

[0016] This step involves constructing a calibration sample dataset that corresponds one-to-one with hyperspectral images and rebound values, so that subsequent steps can train a rebound value mapping model based on hyperspectral images; the operation is as follows: First, each hyperspectral image is corrected one by one, invalid data is removed, and after generating a valid hyperspectral image, the spectral reflectance of each pixel in each band is processed to zero mean and unit variance using the band-based Z-score normalization method to generate a normalized hyperspectral image.

[0017] Secondly, based on the standardized hyperspectral image, a Savitzky-Golay filter is used to smooth and denoise the spectral reflectance vector of each pixel in the hyperspectral image, generating a smoothed and denoised hyperspectral image; and the filtering formula for smoothing and denoising is as follows:

[0018]

[0019] In the formula, For pixels spectral reflectance vector; For pixels The wavelength in the first band is The spectral reflectance value at that location; For pixels The wavelength in the second band is The spectral reflectance value at that location; For pixels In the The wavelength of each band is The spectral reflectance value at that location; This is a transpose operation; for 3D real vector space; For smoothing and noise reduction of pixels p spectral reflectance vector; The Savitzky–Golay filter operator is used to smooth and denoise the pixel spectral reflectance vector, thereby achieving denoising of the standardized hyperspectral image. , These represent the filter window width and the polynomial order, respectively.

[0020] Based on the hyperspectral image after smoothing and denoising, and taking each rebound value measurement point as a reference, the representative rebound value of that rebound value measurement point after angle correction is obtained according to the rebound detection angle correction rule. , For the first The representative rebound value of each rebound measurement point.

[0021] In this step, the rebound test angle correction rule refers to correcting the rebound value according to the provisions of the "Technical Specification for Testing the Compressive Strength of Concrete with Rebound Hammer" (JGJ_T 23-2011), thus obtaining the corrected representative rebound value. (i.e., average rebound value).

[0022] Map the position of the rebound value measurement point in the hyperspectral image to pixel coordinates , , The first The horizontal and vertical coordinates of the corresponding pixels at each rebound value measurement point in the hyperspectral image.

[0023] In pixel coordinates Centered on a predetermined size, a local hyperspectral region is selected. , For the first The local hyperspectral region corresponding to each rebound value measurement point, and the local hyperspectral region The corresponding spectral eigenvector is denoted as ,Right now Indicates the first The spectral feature vector of the local hyperspectral region corresponding to each rebound value measurement point.

[0024] Based on representative rebound values ​​and spectral feature vectors, calibration sample pairs are constructed. Finally, a calibration sample dataset is generated that corresponds one-to-one with hyperspectral images and rebound values.

[0025] In an optional embodiment of the present invention, the preset size is 10×10, that is, the spectral feature vector of a 10×10 pixel area is obtained, and finally the spectral feature vector of the local hyperspectral region corresponding to each rebound value measurement point is used. The input data is the representative rebound value corresponding to the rebound measurement point. To supervise the labeling, several calibration sample pairs were constructed. This allows for the construction of a calibration sample dataset to train a rebound value mapping model based on hyperspectral images.

[0026] Step 3: Using the calibrated sample dataset, train a rebound value mapping model based on hyperspectral images to generate a trained rebound value mapping model based on hyperspectral features, so as to realize the conversion from hyperspectral images to rebound values.

[0027] This step is the process of training a rebound value mapping model based on hyperspectral images, so that subsequent steps can acquire hyperspectral images of the inner wall of the drainage pipe and obtain the rebound value at the corresponding location on the inner wall, thereby realizing the conversion from hyperspectral images to rebound values.

[0028] The structure and connectivity of the rebound value mapping model based on hyperspectral images are as follows: Figure 2 As shown, it includes an encoder and a decoder.

[0029] The encoder includes a high-dimensional feature extraction module and a deep feature extraction module.

[0030] The high-dimensional feature extraction module includes a first convolutional block ConBlock_1, a second convolutional block ConBlock_2, a third convolutional block ConBlock_3, and a first convolutional attention module CBAM_1.

[0031] The deep feature extraction module includes the fourth convolutional block ConBlock_4, the fifth convolutional block ConBlock_5, and the second convolutional attention module CBAM_2.

[0032] Furthermore, the first convolutional block ConBlock_1, the second convolutional block ConBlock_2, the third convolutional block ConBlock_3, the fourth convolutional block ConBlock_4, and the fifth convolutional block ConBlock_5 all include a first convolutional layer, a batch normalization layer, a second ReLU activation function layer, and a regularization layer (Dropout3d).

[0033] Furthermore, both the first and second convolutional attention modules include a channel attention layer and a spatial attention layer; the channel attention layer includes a second global average pooling layer, a first global max pooling layer, a multilayer perceptron, and a first sigmoid activation function layer; the spatial attention layer includes a third global average pooling layer, a second global max pooling layer, a splicing layer, a second convolutional layer, and a second sigmoid activation function layer.

[0034] The decoder includes a first global average pooling layer, a flattening layer, a first fully connected layer, and a second fully connected layer; both the first and second fully connected layers include a Linear layer and a first ReLU activation function layer.

[0035] The process of training a rebound value mapping model based on hyperspectral images using a calibrated sample dataset to generate a trained rebound value mapping model based on hyperspectral images, thereby realizing the conversion from hyperspectral images to rebound values, is as follows: The spectral feature vector of each calibration sample pair in the calibration sample dataset is used as input data, and the representative rebound value of each calibration sample pair is used as output data.

[0036] In this step, the spectral feature vector of each calibration sample pair in the calibration sample dataset is used as input data, and the representative rebound value of each calibration sample pair in the calibration sample dataset is used as output data, which serves as the supervision label. Each input data is a hyperspectral patch of shape (batch, 10, 10, bands), where batch represents the number of samples included in a training batch, and bands represents the number of channels. Finally, after the model is trained, the output data is the predicted rebound value.

[0037] The spectral feature vectors are sequentially input into the first, second, and third convolutional blocks for low- to high-level spatiotemporal spectral feature extraction. After convolution, batch normalization, ReLI activation, and regularization operations, a high-level feature map is generated.

[0038] In this step, after the input data enters the model (a 3D convolutional model), the spectral bands are first shifted to the channel dimension through dimensionality transformation, that is, the shape is transformed from (batch, 10, 10, bands) to (batch, bands, 10, 10), using the number of spectral bands as the number of input channels for the 3D convolution. Subsequently, the spectral feature vector is sequentially input into the first convolutional block ConBlock_1, the second convolutional block ConBlock_2, and the third convolutional block ConBlock_3 for feature extraction, specifically as follows: The first convolutional block expands the number of channels from bands to 64 through 3D convolution (executed within the first convolutional layer), and then performs batch normalization (executed within the batch normalization layer), ReLU activation (executed within the second ReLU activation function layer), and Dropout3d regularization (executed within the regularization layer) to extract low-level spatiotemporal spectral features, outputting a primary feature map with shape (batch, 64, 10, 10). The second convolutional block receives the primary feature map and expands the number of channels from 64 to 128 through the same convolution-normalization-activation-regularization operation, extracting intermediate spatiotemporal spectral features and outputting an intermediate feature map with shape (batch, 128, 10, 10). The third convolutional block receives intermediate feature maps, expanding the number of channels from 128 to 256. Since the time dimension has been compressed to 1, the time dimension of the convolutional kernel of the convolutional layer is reduced to 1, focusing on the joint extraction of spatial and spectral features, and outputting a high-level feature map with the shape (batch, 256, 10, 10).

[0039] The high-level feature map is input into the first convolutional attention module, where channel and spatial attention are weighted and fused to generate the first fused feature map.

[0040] In this step, the high-level feature map output by the third convolutional block ConBlock_3 is input into the first convolutional attention module CBAM_1 for dual attention weighting of channels and space, specifically: First, channel attention operations are performed. Global average pooling (executed in the second global average pooling layer) and global max pooling (executed in the first global max pooling layer) are applied to the high-level feature map, compressing it to (batch, 256, 1, 1). Then, it is processed by a multilayer perceptron (MLP) with a compression ratio of 8 (256→32→256) to extract the inter-channel dependencies. The two results are then added together and passed through the first sigmoid activation function layer to generate channel attention weights. Finally, the channel attention weights are multiplied with the high-level feature map to generate the output feature map F' of the channel attention layer, thereby strengthening the spectral channels and suppressing redundant channels.

[0041] Then, spatial attention operations are performed: the input to the spatial attention module is the output feature map F' of the channel attention module, with a size of (batch, 256, 10, 10). Here, the channel dimension refers to the 256 feature channels in the output feature map F'. To generate spatial attention weights, average pooling (performed within the third global average pooling layer) and max pooling (performed within the second global max pooling layer) are performed along the channel dimension of F', respectively, resulting in two spatial description maps with a size of (batch, 1, 10, 10). These two maps are then concatenated along the channel dimension to form a feature map with a size of (batch, 2, 10, 10), and spatial attention weights are generated through a 7×7 convolutional layer (performed within the second convolutional layer) and a sigmoid activation function (performed within the second sigmoid activation function layer). Finally, this spatial attention weight is multiplied element-wise with the output feature map F' of the channel attention module to obtain the first fused feature map F'', which is a refined feature map after dual attention weighting. Since the spatial attention weights are applied to all 256 channels via a broadcast mechanism in the channel dimension, the size of the first fused feature map F'' remains (batch, 256, 10, 10).

[0042] The first fused feature map is sequentially input into the fourth and fifth convolutional blocks for deep feature extraction. After convolution, batch normalization, ReLI activation and regularization operations, a deep feature map is generated.

[0043] In this step, the first fused feature map is sequentially input into the fourth convolutional block ConBlock_4 and the fifth convolutional block ConBlock_5 for deep feature extraction. The fourth convolutional block expands the number of channels from 256 to 512, extracting deeper discriminative features based on the attention-refined features, and outputs a feature map with the shape (batch, 512, 10, 10). The fifth convolutional block keeps the number of channels unchanged at 512 and further refines the feature map output by the fourth convolutional block, outputting a refined deep feature map with the same shape (batch, 512, 10, 10).

[0044] The deep feature map is input into the second convolutional attention module, where channel and spatial attention are weighted and fused to generate the second fused feature map.

[0045] In this step, the deep feature map output from the fifth convolutional block is input into the second convolutional attention module CBAM_2 for dual attention weighting of the deep features. The operation process is the same as the generation process of the first fused feature map, and will not be repeated here. The difference is that it is applied to 512 channels. The channel attention MLP compression structure is 512→64→512. Therefore, this step re-weights and refines the channels and spatial positions of the deep features, and outputs a final refined feature map with the shape (batch, 512, 10, 10), which is the second fused feature map.

[0046] The second fused feature map is sequentially fed into the first global average pooling layer, the flattening layer, the first fully connected layer, and the second fully connected layer. After global averaging and flattening, a linear mapping is performed to generate the rebound value prediction result. In this step, the second fused feature map is input into the first global average pooling layer, Global_pool, where global averaging is performed on the time, height, and width dimensions to eliminate spatial differences. The output shape is (batch, 512, 1, 1), and after flattening (operation is performed on the flattening layer), a feature vector of (batch, 512) is obtained. Finally, this feature vector is sequentially input into the Linear layer of the first fully connected layer (512→256), the first ReLU activation function layer, and then undergoes linear transformation and ReLU activation. It is then passed through the Linear layer of the second fully connected layer (256→1), the first ReLU activation function layer, and undergoes linear transformation and ReLU activation again, mapping it to a single continuous value, which is the model's prediction result for the concrete rebound value.

[0047] A cross-entropy loss function is introduced to backpropagate the model parameters of the rebound value mapping model based on hyperspectral images, and finally a trained rebound value mapping model based on hyperspectral images is generated.

[0048] In this step, during model training, a cross-entropy loss function is introduced to optimize and adjust the model parameters, ultimately generating a trained rebound value mapping model based on hyperspectral images. Simultaneously, during training, the smoothing coefficient is set to 0.1 to effectively alleviate overfitting of the model to a single class label. Furthermore, the AdamW optimizer is used for parameter updates, with regularization constraints implemented through weight decay (weight_decay=0.01). Additionally, during training, a cosine annealing learning rate scheduler is used to restart the learning rate scheduling strategy (CosineAnnealingWarmRestarts), setting the initial period length to 10, the period rate to 2, and the minimum learning rate to 1e. -6 The training iterations are set to 100. During training, the training loss, validation loss, training accuracy, and validation accuracy are recorded simultaneously to determine the convergence state. Finally, the coefficient of determination R is used. 2Evaluation metrics such as mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) are used to assess the learning performance of the rebound value mapping model.

[0049] The formula for calculating the cross-entropy loss function is as follows:

[0050] in, The cross-entropy loss function value is used to measure the true probability distribution. With the predicted probability distribution The degree of difference between them; For the first Sample The probability distribution in real-world events; It is a logarithmic function; For the first Sample The probability distribution in the predicted event.

[0051] Step 4: Acquire hyperspectral images of the inner wall surface of the drainage pipe to be tested, perform data preprocessing, and input them into the trained rebound value mapping model based on hyperspectral images to generate rebound values ​​at corresponding locations on the inner wall of the drainage pipe.

[0052] In this step, a hyperspectral camera is mounted on an existing elastic wave pipeline inspection robot structure to acquire hyperspectral images of the inner wall surface of the drainage pipe under test. Figure 3 As shown, it includes a hyperspectral camera 1, a front-end searchlight and projection camera group 2, and an elastic wave pipeline inspection robot; wherein, the elastic wave pipeline inspection robot includes a main load-bearing structure 3, a scissor lift 4, an all-terrain propeller 5, and a power and transmission system module 6.

[0053] The hyperspectral camera, integrating a halogen light source, is mounted on the main support structure 3. It collects hyperspectral data from the surface of the target object, enabling non-contact spectral acquisition of concrete surface information and providing a data foundation for subsequent mechanical property inversion and defect identification. A front-end searchlight and projection camera group acquire visible light video information and provide supplementary lighting. A full-body protective bracket and cover provide physical protection for the machine body and key sensors in complex working conditions, reducing the risk of collision damage during operation. The main support structure, including the upper platform, side beams, and bottom frame, can be manufactured using lightweight aluminum alloy materials to balance structural strength and overall machine weight. In the example, an elastic wave detection sensor is mounted on the upper part of the platform for auxiliary calibration. A scissor lift allows for height adjustment of the detection platform, dynamically adjusting the sensor installation height according to different detection distances and working conditions. An all-terrain propeller ensures stable robot movement in complex environments, adapting to movement needs inside pipelines, on potholes, and in uneven working areas. The power and transmission system module, including components such as hub motors, gearboxes, and couplings, is used to drive the wheels, the lifting mechanism, and the attitude adjustment and rotation control of the hyperspectral detection platform, thereby achieving coordinated linkage and precise drive of multiple actuators.

[0054] Step 5: Based on the rebound values ​​at corresponding locations on the inner wall of the drainage pipe, and combining the conversion relationship between rebound values ​​and compressive strength, calculate the concrete compressive strength, and generate the spatial distribution results of the compressive strength of the inner wall of the drainage pipe, specifically: First, based on the rebound value of the corresponding location on the inner wall of the drainage pipe, and combined with the conversion relationship between the rebound value and compressive strength, the compressive strength of the concrete is calculated, and the compressive strength matrix of the inner wall of the drainage pipe is generated.

[0055] In this step, the formula for calculating the compressive strength of concrete is:

[0056]

[0057] in, This is the average rebound value after angle correction; This is the original average rebound value; For testing angles; This is the angle correction value; It refers to the compressive strength of concrete; As the benchmark coefficient; The index is the sensitivity index for rebound value; The coefficient representing the influence of carbonization depth; This refers to the carbonization depth.

[0058] The parameters in the above formula (such as the rebound value sensitivity index) The values ​​of (etc.) are taken with reference to the Technical Specification for Testing the Compressive Strength of Concrete with Rebound Hammer (JGJ_T 23-2011).

[0059] Secondly, the compressive strength matrix of the inner wall of the drainage pipe is standardized and structured to generate compressive strength matrix data in two-dimensional or three-dimensional matrix form.

[0060] Then, based on the coordinate relationship between the inner wall of the drainage pipe or the circumference and the axis, the compressive strength matrix data is spatially mapped to construct a two-dimensional grid coordinate system consistent with the geometry of the inner wall of the drainage pipe.

[0061] Finally, the compressive strength matrix data is normalized or standardized, color mapping rules are set, and the processed compressive strength matrix data is rendered as a heat map of the compressive strength distribution of the inner wall of the drainage pipe, thus obtaining the spatial distribution result of the compressive strength of the inner wall of the drainage pipe.

[0062] In this step, the compressive strength matrix data will be normalized or standardized, and color mapping rules will be set. Based on the Matplotlib, Seaborn, or Plotly visualization library, the corresponding heatmap drawing function will be called to render the processed compressive strength matrix data into a heatmap of the compressive strength distribution of the inner wall of the drainage pipe. Then, it will be exported and stored as an image file or interactive visualization file format for subsequent analysis, report presentation, or engineering decision support.

[0063] A device for acquiring hyperspectral images includes a hyperspectral camera 1, a front-end searchlight and a projection camera group 2, and an elastic wave pipeline inspection robot.

[0064] The elastic wave pipeline inspection robot includes a main load-bearing structure 3, a scissor lift 4, an all-terrain propeller 5, and a power and transmission system module 6.

[0065] The hyperspectral camera, integrating a halogen light source, is mounted on the main support structure 3. It collects hyperspectral data from the surface of the target object, enabling non-contact spectral acquisition of concrete surface information and providing a data foundation for subsequent mechanical property inversion and defect identification. A front-end searchlight and projection camera group acquire visible light video information and provide supplementary lighting. A full-body protective bracket and cover provide physical protection for the machine body and key sensors in complex working conditions, reducing the risk of collision damage during operation. The main support structure, including the upper platform, side beams, and bottom frame, can be manufactured using lightweight aluminum alloy materials to balance structural strength and overall machine weight. In the example, an elastic wave detection sensor is mounted on the upper part of the platform for auxiliary calibration. A scissor lift allows for height adjustment of the detection platform, dynamically adjusting the sensor installation height according to different detection distances and working conditions. An all-terrain propeller ensures stable robot movement in complex environments, adapting to movement needs inside pipelines, on potholes, and in uneven working areas. The power and transmission system module, including components such as hub motors, gearboxes, and couplings, is used to drive the wheels, the lifting mechanism, and the attitude adjustment and rotation control of the hyperspectral detection platform, thereby achieving coordinated linkage and precise drive of multiple actuators.

[0066] In summary, the pipeline compressive strength assessment method and equipment proposed in this invention, which uses hyperspectral data processing and rebound value correlation calibration, introduces calibration samples to train and correct the parameters of the rebound value mapping model, constructs a rebound value mapping model based on hyperspectral images, and generates rebound values ​​at corresponding locations on the inner wall of the drainage pipeline. This allows for the calculation of concrete compressive strength, generating a stable and reliable spatial distribution result of the inner wall compressive strength of the drainage pipeline, thus achieving integrated processing of hyperspectral image input and compressive strength parameter output. Furthermore, this invention significantly improves the automation level and measurement efficiency of drainage pipeline compressive strength testing, reduces manual intervention and on-site operation risks, and provides reliable technical support for drainage pipeline structural safety assessment, operation and maintenance, and risk early warning.

[0067] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0068] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for evaluating the compressive strength of pipes based on hyperspectral correlation and resilience value calibration, characterized in that, Includes the following steps: Rebound value measuring points were set up in the accessible area of ​​the drainage pipe to collect rebound value data, and hyperspectral images containing the rebound value measuring points were also collected. After preprocessing the hyperspectral images, a calibration sample dataset is constructed by combining the rebound value data. Using a calibrated sample dataset, a rebound value mapping model based on hyperspectral images is trained to generate a trained rebound value mapping model based on hyperspectral images, thereby realizing the conversion from hyperspectral images to rebound values. Hyperspectral images of the inner wall surface of the drainage pipe to be tested are acquired. After data preprocessing, the images are input into a trained rebound value mapping model based on hyperspectral images to generate rebound values ​​at corresponding locations in the inner wall region of the drainage pipe. Based on the rebound values ​​at corresponding locations on the inner wall of the drainage pipe, and combined with the conversion relationship between rebound values ​​and compressive strength, the compressive strength of concrete is calculated, generating the spatial distribution results of the compressive strength of the inner wall of the drainage pipe.

2. The method for pipe compressive strength evaluation by hyperspectral and rebound value correlation calibration according to claim 1, characterized in that, The process of constructing a calibration sample dataset by preprocessing hyperspectral images and combining them with rebound value data is as follows: After correcting each hyperspectral image, removing invalid data, and generating valid hyperspectral images, standardization processing is performed to generate standardized hyperspectral images. Based on the standardized hyperspectral image, the spectral reflectance vector of each pixel in the hyperspectral image is smoothed and denoised to generate a smoothed and denoised hyperspectral image. Based on the hyperspectral image after smoothing and denoising, and taking each rebound value measurement point as a reference, the representative rebound value of that rebound value measurement point after angle correction is obtained according to the rebound detection angle correction rule. , For the first Representative rebound values ​​of each rebound measurement point; mapping the location of the rebound value measurement point in the hyperspectral image to pixel coordinates , , are the horizontal and vertical coordinates of the pixel corresponding to the first th rebound value measurement point in the hyperspectral image, respectively In pixel coordinates , a local hyperspectral region of a preset size is intercepted , , a local hyperspectral region corresponding to the first elasticity value measurement point, and a spectral feature vector corresponding to the local hyperspectral region is denoted as ; Based on the representative rebound value, the spectral feature vector, the calibration sample pair is constructed Finally, the hyperspectral image and the calibration sample dataset corresponding to the rebound value are generated.

3. The method for evaluating the compressive strength of pipelines based on hyperspectral and resilience value correlation calibration according to claim 2, characterized in that, The filtering formula for smoothing and denoising the spectral reflectance vector of each pixel in the hyperspectral image is as follows: in, For pixels spectral reflectance vector; For pixels The wavelength in the first band is The spectral reflectance value at that location; For pixels The wavelength in the second band is The spectral reflectance value at that location; For pixels In the The wavelength of each band is The spectral reflectance value at that location; This is a transpose operation; for 3D real vector space; For smoothing and noise reduction of pixels p spectral reflectance vector; For the Savitzky–Golay filter operator; , These represent the filter window width and the polynomial order, respectively.

4. The method for pipe buckling strength evaluation by hyperspectral and rebound value correlation calibration according to claim 1, characterized in that, The rebound value mapping model based on hyperspectral images includes an encoder and a decoder; The encoder includes a high-dimensional feature extraction module and a deep feature extraction module; The high-dimensional feature extraction module includes a first convolutional block, a second convolutional block, a third convolutional block, and a first convolutional attention module; The deep feature extraction module includes a fourth convolutional block, a fifth convolutional block, and a second convolutional attention module; The decoder includes a first global average pooling layer, a flattening layer, a first fully connected layer, and a second fully connected layer; both the first and second fully connected layers include a Linear layer and a first ReLU activation function layer.

5. The method for pipe buckling strength evaluation by hyperspectral and rebound value correlation calibration according to claim 4, characterized in that, The first, second, third, fourth, and fifth convolutional blocks each include a first convolutional layer, a batch normalization layer, a second ReLU activation function layer, and a regularization layer.

6. The method for pipe buckling strength evaluation by hyperspectral and rebound value correlation calibration according to claim 4, characterized in that, Both the first convolutional attention module and the second convolutional attention module include a channel attention layer and a spatial attention layer; The channel attention layer includes a second global average pooling layer, a first global max pooling layer, a multilayer perceptron, and a first sigmoid activation function layer; The spatial attention layer includes a third global average pooling layer, a second global max pooling layer, a splicing layer, a second convolutional layer, and a second sigmoid activation function layer.

7. The method according to claim 5 or 6, wherein, Using a calibrated sample dataset, a rebound value mapping model based on hyperspectral images is trained to generate a trained rebound value mapping model based on hyperspectral images. The process of converting hyperspectral images to rebound values ​​is as follows: The spectral feature vector of each calibration sample pair in the calibration sample dataset is used as input data, and the representative rebound value of each calibration sample pair is used as output data. The spectral feature vectors are sequentially input into the first convolutional block, the second convolutional block, and the third convolutional block to extract spatiotemporal spectral features from low to high level. After convolution, batch normalization, ReLU activation, and regularization operations, a high-level feature map is generated. The high-level feature map is input into the first convolutional attention module, where channel and spatial attention are weighted and fused to generate the first fused feature map. The first fused feature map is sequentially input into the fourth and fifth convolutional blocks for deep feature extraction. After convolution, batch normalization, ReLI activation and regularization operations, a deep feature map is generated. The deep feature map is input into the second convolutional attention module, where channel and spatial attention are weighted and fused to generate the second fused feature map. The second fused feature map is sequentially fed into the first global average pooling layer, the flattening layer, the first fully connected layer, and the second fully connected layer. After global averaging and flattening, a linear mapping is performed to generate the rebound value prediction result. A cross-entropy loss function is introduced to backpropagate the model parameters of the rebound value mapping model based on hyperspectral images, and finally a trained rebound value mapping model based on hyperspectral images is generated.

8. The method for pipe buckling strength evaluation by hyperspectral and rebound value correlation calibration according to claim 1, characterized in that, The process of calculating the concrete compressive strength and generating the spatial distribution results of the inner wall compressive strength of the drainage pipe based on the rebound values ​​at corresponding locations on the inner wall of the drainage pipe, combined with the conversion relationship between rebound values ​​and compressive strength, is as follows: Based on the rebound value of the corresponding location on the inner wall of the drainage pipe, and combined with the conversion relationship between the rebound value and the compressive strength, the compressive strength of the concrete is calculated, and the compressive strength matrix of the inner wall of the drainage pipe is generated. The compressive strength matrix of the inner wall of the drainage pipe is standardized and structured to generate compressive strength matrix data in two-dimensional or three-dimensional matrix form. Based on the coordinate relationship between the inner wall of the drainage pipe or the circumference and the axis, spatial mapping is performed on the compressive strength matrix data to construct a two-dimensional grid coordinate system consistent with the geometry of the inner wall of the drainage pipe. Normalize or standardize the compressive strength matrix data, set color mapping rules, and render the processed compressive strength matrix data as a heat map of the compressive strength distribution of the inner wall of the drainage pipe, thus obtaining the spatial distribution result of the compressive strength of the inner wall of the drainage pipe.

9. The method for evaluating the compressive strength of pipelines based on hyperspectral and resilience value correlation calibration according to claim 8, characterized in that, The formula for calculating the compressive strength of concrete is: in, This is the average rebound value after angle correction. This is the original average rebound value. For testing angles, This is the angle correction value. For concrete compressive strength, As the benchmark coefficient, The index is the sensitivity table for rebound value. The coefficient representing the influence of carbonization depth. This refers to the carbonization depth.

10. An apparatus for acquiring hyperspectral images, characterized by It includes a hyperspectral camera (1), a front-end searchlight and a projection camera group (2), and an elastic wave pipeline inspection robot.

Images