Construction concealed defect detection method and system

Abstract

This invention provides a method and system for detecting hidden defects in construction, including acquiring raw data of the structure to be detected through multi-channel sensors, generating structural feature data containing material properties through preprocessing, constructing a state transition matrix based on material property gradients and geometric distances, and performing spatial state estimation by combining an observation matrix matched with material categories to obtain the defect probability state value and covariance of each sampling point. The probability state value is used as a feature map input to a detection network model using a piecewise Swish activation function to achieve defect identification. The piecewise threshold of the activation function is determined based on the statistical distribution of the feature map and the state estimation covariance, and the scaling factor of each interval is adjusted according to the interval width and input statistical characteristics, thereby improving the defect detection accuracy.

Description

Technical Field

[0001] This application belongs to the field of testing, and in particular relates to a method and system for detecting hidden defects in construction. Background Technology

[0002] Traditional methods for detecting hidden defects mainly employ non-destructive testing techniques, such as ground-penetrating radar, ultrasonic testing, and infrared thermal imaging, which infer the internal condition of a structure by analyzing signals collected by sensors. However, the complex environment of construction sites, including the non-uniformity of structural materials, humidity variations, and interference from the distribution of reinforcing bars, can lead to low signal-to-noise ratios and blurred signal characteristics in the raw sensor data. In the practical application of deep learning models for defect detection, the choice of activation function plays a decisive role in the network's performance. Currently, most defect detection networks use general activation functions, such as ReLU and its variants, Sigmoid, or the standard Swish function. For hidden defect detection tasks, the defect feature maps obtained after preliminary processing often have a unique, non-uniform statistical distribution. Most areas are healthy background regions with values ​​concentrated in a certain range, while a few abnormal defect regions exhibit different numerical characteristics. For example, it may suppress most background areas, causing the loss of weak defect feature information during propagation; or it may be sensitive to noise, misclassifying some high-uncertainty background areas as defects. Current technologies cannot feed back and integrate the feature uncertainty information obtained in the preprocessing stages into the structural planning of subsequent deep learning networks, especially to guide the behavior of activation functions. Therefore, how to plan a detection method that can be adjusted according to the statistical properties and uncertainties of the input feature map is a technical challenge that urgently needs to be solved in the field. Summary of the Invention

[0003] To address the aforementioned problems, in a first aspect of this invention, a method for detecting hidden defects in construction is proposed, comprising the following steps: Acquire raw data from multi-channel sensors of the construction structure to be inspected, and preprocess it to generate structural feature data containing material properties of each spatial sampling point; A spatial state model is constructed, and the state of the structural feature data is estimated to obtain the hidden defect probability state value and the corresponding state estimation covariance of each spatial sampling point. The state transition process of the model determines the state transition matrix based on the material property gradient and geometric distance between the current sampling point and other sampling points in the spatial neighborhood. The observation process of the model matches different observation matrices for the sampling points according to the preset material category and updates the state in combination with the structural feature data. The hidden defect probability state values ​​obtained above are used as feature maps and input into a defect detection network model employing a segmented Swish activation function to obtain defect detection results. The segment threshold of the segmented Swish activation function is determined based on the statistical distribution of the feature map and the corresponding state estimation covariance. Furthermore, for input values ​​falling into different segment intervals defined by the segment threshold, a scaling factor corresponding to the width of the segment interval and the statistical characteristics of the input values ​​falling into the interval is used for calculation.

[0004] Optionally, the step of acquiring the raw data from the multi-channel sensors of the construction structure to be inspected and preprocessing it to generate structural feature data containing the material properties of each spatial sampling point includes: A three-dimensional Gaussian filter is used to smooth and denoise the raw data from the multi-channel sensors. The denoised data is subjected to min-max normalization to linearly map all data values ​​to the interval [0,1]. Furthermore, the values ​​of density, elastic modulus, and Poisson's ratio are extracted from the normalized data for each spatial sampling point as material properties, which together constitute the structural feature data.

[0005] Optionally, the state transition process of the model determines the state transition matrix based on the material property gradient and geometric distance between the current sampling point and other sampling points in the spatial neighborhood, including: For the current sampling point i, select the k sampling points with the closest three-dimensional Euclidean distance to the point as the spatial neighborhood; Calculate the material property vector between the current sampling point i and any sampling point j in the neighborhood. and square Euclidean distance and the square of the geometric distance Before calculation, the three-dimensional geometric coordinates of the sampling points are subjected to min-max normalization to make... It becomes a dimensionless quantity; According to the formula Calculate the state transition elements, and all calculated elements. Together they form the state transition matrix, where k is a positive integer. and These are the preset weighting coefficients.

[0006] Optionally, the observation process of the model matches different observation matrices for sampling points according to a preset material category, including: The materials are pre-categorized into concrete, reinforced steel, and voids; When the current sampling point is determined to belong to concrete based on the structural feature data, an observation matrix is ​​matched for that point. ; When the point is identified as a reinforcing bar, the observation matrix is ​​matched. ; When a hole is identified, the observation matrix is ​​matched for that point. ,in , , , , and These are the preset observation parameters corresponding to each material category.

[0007] Optionally, the segmentation threshold of the segmented Swish activation function is determined based on the statistical distribution of the feature map and the corresponding state estimation covariance, including: Calculate the mean of the probability state values ​​of all hidden defects in the feature map. and standard deviation ; Calculate the average trace of the state estimation covariance matrix at the corresponding location in the feature map. ; And according to the formula and Set two segment thresholds and ,in This is the preset adjustment coefficient.

[0008] Optionally, the step of calculating the scaling factor for input values ​​falling into different segment intervals divided by the segmentation threshold, using a scaling factor corresponding to the width of the segment interval and the statistical characteristics of the input values ​​falling into the interval, includes: For input value x, when When using a scaling factor ; when When using a scaling factor ; when When using a scaling factor ; And the input value x and the corresponding scaling factor Substitute into Swish activation function The calculations are performed in the middle.

[0009] Optionally, the defect detection network model employing the piecewise Swish activation function includes: An encoder path consisting of N convolutional layers and M max pooling layers is used to extract features step by step; A decoder path consisting of N convolutional layers and M upsampling layers is used to progressively restore the feature map resolution. Skip connections are set between corresponding layers of the encoder and decoder paths to fuse shallow and deep features; Furthermore, all convolutional layers in the encoder and decoder paths employ the segmented Swish activation function, where N and M are preset positive integers.

[0010] In a second aspect of the invention, a construction concealed defect detection system is provided, comprising the following modules: The generation module is used to acquire the raw data from the multi-channel sensors of the construction structure to be inspected, and to preprocess the data to generate structural feature data containing the material properties of each spatial sampling point. A construction module is used to construct a spatial state model, perform state estimation on the structural feature data, and obtain the hidden defect probability state value and the corresponding state estimation covariance of each spatial sampling point. The state transition process of the model determines the state transition matrix based on the material property gradient and geometric distance between the current sampling point and other sampling points in the spatial neighborhood. The observation process of the model matches different observation matrices for the sampling points according to a preset material category and updates the state in conjunction with the structural feature data. The calculation module is used to input the obtained hidden defect probability state value as a feature map into a defect detection network model using a piecewise Swish activation function to obtain defect detection results; wherein, the segment threshold of the piecewise Swish activation function is determined based on the statistical distribution of the feature map and the corresponding state estimation covariance; and, for input values ​​falling into different segment intervals divided by the segment threshold, a scaling factor corresponding to the width of the segment interval and the statistical characteristics of the input values ​​falling into the interval is used for calculation.

[0011] Preferably, the step of acquiring the raw data from the multi-channel sensors of the construction structure to be inspected, and preprocessing it to generate structural feature data containing the material properties of each spatial sampling point, includes: A three-dimensional Gaussian filter is used to smooth and denoise the raw data from the multi-channel sensors. The denoised data is subjected to min-max normalization to linearly map all data values ​​to the interval [0,1]. Furthermore, the values ​​of density, elastic modulus, and Poisson's ratio are extracted from the normalized data for each spatial sampling point as material properties, which together constitute the structural feature data.

[0012] Preferably, the state transition process of the model determines the state transition matrix based on the material property gradient and geometric distance between the current sampling point and other sampling points in the spatial neighborhood, including: For the current sampling point i, select the k sampling points with the closest three-dimensional Euclidean distance to the point as the spatial neighborhood; Calculate the material property vector between the current sampling point i and any sampling point j in the neighborhood. and square Euclidean distance and the square of the geometric distance Before calculation, the three-dimensional geometric coordinates of the sampling points are subjected to min-max normalization to make... It becomes a dimensionless quantity; According to the formula Calculate the state transition elements, and all calculated elements. Together they form the state transition matrix, where k is a positive integer. and These are the preset weighting coefficients.

[0013] Preferably, the observation process of the model matches different observation matrices for sampling points according to a preset material category, including: The materials are pre-categorized into concrete, reinforced steel, and voids; When the current sampling point is determined to belong to concrete based on the structural feature data, an observation matrix is ​​matched for that point. ; When the point is identified as a reinforcing bar, the observation matrix is ​​matched. ; When a hole is identified, the observation matrix is ​​matched for that point. ,in , , , , and These are the preset observation parameters corresponding to each material category.

[0014] Preferably, the segmentation threshold of the segmented Swish activation function is determined based on the statistical distribution of the feature map and the corresponding state estimation covariance, including: Calculate the mean of the probability state values ​​of all hidden defects in the feature map. and standard deviation ; Calculate the average trace of the state estimation covariance matrix at the corresponding location in the feature map. ; And according to the formula and Set two segment thresholds and ,in This is the preset adjustment coefficient.

[0015] Preferably, the step of calculating the scaling factor for input values ​​falling into different segment intervals divided by the segmentation threshold, using a scaling factor corresponding to the width of the segment interval and the statistical characteristics of the input values ​​falling into the interval, includes: For input value x, when When using a scaling factor ; when When using a scaling factor ; when When using a scaling factor ; And the input value x and the corresponding scaling factor Substitute into Swish activation function The calculations are performed in the middle.

[0016] Preferably, the defect detection network model employing the piecewise Swish activation function includes: An encoder path consisting of N convolutional layers and M max pooling layers is used to extract features step by step; A decoder path consisting of N convolutional layers and M upsampling layers is used to progressively restore the feature map resolution. Skip connections are set between corresponding layers of the encoder and decoder paths to fuse shallow and deep features; Furthermore, all convolutional layers in the encoder and decoder paths employ the segmented Swish activation function, where N and M are preset positive integers.

[0017] This invention constructs a spatial state model that considers the spatial neighborhood relationships and material property gradients between sampling points when processing multi-channel sensor data. It also matches different observation processes to different material categories, thereby suppressing noise and interference from material inhomogeneities in the original data. The probability state values ​​and their corresponding estimated covariances are used to guide the construction of piecewise activation functions in the defect detection network. This allows the activation functions to handle features in different numerical ranges, suppressing highly deterministic background regions and focusing on regions that may represent weak defects but have uncertainty. This enhances the model's ability to identify weak defect features and improves the accuracy and reliability of detecting hidden defects in construction. Attached Figure Description

[0018] Figure 1 A flowchart of the first embodiment; Figure 2 This is a schematic diagram of the observation matrix matching mechanism for heterogeneous materials. Figure 3 This is a schematic diagram illustrating the principle of the segmented Swish activation function; Figure 4 This is a schematic diagram of threshold division based on the probability distribution of hidden defects. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In the first embodiment, the present invention proposes a method for detecting hidden defects in construction, such as... Figure 1 This includes the following steps: S1. Acquire the raw data from the multi-channel sensors of the construction structure to be inspected, and perform preprocessing to generate structural feature data containing the material properties of each spatial sampling point; Taking ground-penetrating radar (GPR) detection of concrete structures as an example, the GPR equipment scans the target structure to acquire radar echo signals from each spatial sampling point on a three-dimensional grid, i.e., the raw data from the multi-channel sensor. The raw data undergoes preprocessing, including zero-point correction to align the signal start time, signal gain to compensate for deep signal attenuation, and background removal to eliminate static interference such as surface reflection. After preprocessing, a Hilbert transform is performed on the echo signal from each sampling point to extract multiple features such as instantaneous amplitude, instantaneous phase, and instantaneous frequency. These features reflect changes in the dielectric constant of the medium and can represent the properties of different materials such as concrete, reinforcing steel, and voids. The multiple features extracted from each sampling point are combined into a feature vector, and the feature vectors of all sampling points constitute structural feature data covering the entire detection area.

[0021] In an optional embodiment, the step of acquiring the raw multi-channel sensor data of the construction structure to be inspected and preprocessing it to generate structural feature data containing material properties of each spatial sampling point includes: A three-dimensional Gaussian filter is used to smooth and denoise the raw data from the multi-channel sensors. The denoised data is subjected to min-max normalization to linearly map all data values ​​to the interval [0,1]. Furthermore, the values ​​of density, elastic modulus, and Poisson's ratio are extracted from the normalized data for each spatial sampling point as material properties, which together constitute the structural feature data.

[0022] The preprocessing process receives the raw 3D data volume from the ground-penetrating radar, which can be viewed as a multi-channel 3D matrix, such as a volumetric data with dimensions of 256×256×256, where each spatial sampling point contains measurements from multiple channels. To eliminate sensor noise and measurement errors, a 3×3×3 3D Gaussian filter is used to convolve the data volume, achieving smooth denoising while preserving important structural information such as material boundaries.

[0023] After denoising, the data is subjected to min-max normalization. From the normalized multi-channel data, three key material properties are extracted for each spatial sampling point: density, elastic modulus, and Poisson's ratio. For example, based on the radar wave propagation velocity formula... ,in The relative permittivity, The vacuum permittivity, The density, elastic modulus, and Poisson's ratio are obtained by inverting the vacuum permeability, combined with the normalized echo amplitude and propagation time, through a pre-trained BP neural network model. The BP neural network model is a three-layer fully connected structure. The input layer dimension corresponds to the dimension of the preprocessed multi-channel signal feature vector, such as the instantaneous amplitude, instantaneous phase, and instantaneous frequency three-dimensional features extracted from ground-penetrating radar signals via Hilbert transform. There are two hidden layers: the first layer contains 128 neurons, and the second layer contains 64 neurons, both using the ReLU activation function to enhance nonlinear fitting ability. The output layer contains three neurons, corresponding to the normalized predicted values ​​of density, elastic modulus, and Poisson's ratio, respectively. The training process uses standard material samples, such as signal features of concrete, steel reinforcement, and void standard specimens with known density, elastic modulus, and Poisson's ratio as input, and real material properties as labels to construct training and validation sets. The sample ratio is 7:3. The loss function uses mean squared error to minimize the deviation between predicted and true attributes. The optimizer is Adam, with a learning rate of 1e-4 and a batch size of 32. During training, the predicted output is calculated through forward propagation, and the network weights and biases are updated layer by layer using the backpropagation algorithm. The training is iterated until the validation set loss does not decrease for 10 consecutive rounds or reaches the maximum number of iterations of 500 rounds, resulting in a stable signal material property inversion model, which ensures that the three target material properties are accurately mapped from the preprocessed signal features.

[0024] A three-dimensional feature vector is constructed using density, elastic modulus, and Poisson's ratio. For example, the structural feature data of a sampling point located at coordinates x, y, z can be represented as a vector [0.8, 0.7, 0.3], corresponding to the normalized density, elastic modulus, and Poisson's ratio values, respectively. The set of feature vectors for all sampling points constitutes the structural feature data.

[0025] S2, Construct a spatial state model, perform state estimation on the structural feature data, and obtain the hidden defect probability state value and the corresponding state estimation covariance of each spatial sampling point; wherein, the state transition process of the model determines the state transition matrix based on the material property gradient and geometric distance between the current sampling point and other sampling points in the spatial neighborhood; the observation process of the model matches different observation matrices for the sampling points according to the preset material category, and updates the state in combination with the structural feature data; Optionally, the spatial state model is constructed by discretizing the construction structure to be inspected into several spatial sampling points. The probability of hidden defects and its related auxiliary quantities at each sampling point are used as state variables. A state transition relationship is established based on the material property differences and geometric distances between adjacent sampling points, thereby obtaining a state transition matrix that reflects spatial correlation. At the same time, the structural feature data extracted by multi-channel sensors are used as observations. The corresponding observation matrix is ​​matched according to the material category of the sampling point to realize the mapping between state and observation. On this basis, a state estimation method is introduced to jointly estimate the state and uncertainty of each sampling point, thereby obtaining a spatial state model that describes the spatial distribution of hidden defects in the construction structure.

[0026] The hidden defect probability of each spatial sampling point is defined as the state of that point. The structural feature data obtained above is initially classified using methods such as K-means clustering, dividing all sampling points into preset material categories such as concrete, reinforcing steel, and potential anomaly zones. For the state transition process, taking any sampling point as the center, 26 neighboring sampling points are selected. The Euclidean distance of the structural feature data between the center point and each neighboring point is calculated as the material property gradient, and their spatial geometric distance is also calculated. In the state transition matrix, the transition probability of a neighboring point to the center point is set as a function inversely proportional to both the material property gradient and the geometric distance, such as a Gaussian function. Each row of the state transition matrix is ​​normalized; the more similar the materials and the closer the distance between the neighboring points and the center point, the greater the influence of the state on the prediction of the center point's state. For the observation process, an independent observation matrix is ​​defined for each preset material category. This matrix represents the mapping relationship between the actual defect probability state and the observed structural feature data under that material category. For example, for the healthy concrete category, the observation matrix reflects that a low defect probability corresponds to normal concrete feature data.

[0027] During state estimation, the prior state of the current point is predicted using the states of neighboring points based on the state transition matrix. The corresponding observation matrix is ​​selected based on the pre-classified material category of the current point, and structural feature data is used as the observation input. The residual between the observed value and the predicted prior state value is calculated using the observation matrix. The Kalman gain is calculated using the residual information, prior covariance, and observation noise covariance. The prior state is then weighted and corrected using the Kalman gain to obtain the hidden defect probability state value and the updated state estimation covariance. The covariance represents the degree of uncertainty in this probability estimation.

[0028] In an optional embodiment, the state transition process of the model determines the state transition matrix based on the material property gradient and geometric distance between the current sampling point and other sampling points in the spatial neighborhood, including: For the current sampling point i, select the k sampling points with the closest three-dimensional Euclidean distance to the point as the spatial neighborhood; Calculate the material property vector between the current sampling point i and any sampling point j in the neighborhood. and square Euclidean distance and the square of the geometric distance Before calculation, the three-dimensional geometric coordinates of the sampling points are subjected to min-max normalization to make... It becomes a dimensionless quantity; According to the formula Calculate the state transition elements, and all calculated elements. Together, they constitute the state transition matrix. Each row of the state transition matrix is ​​then normalized. where k is a positive integer. and The preset weighting coefficients can be calibrated using cross-validation.

[0029] Let the neighborhood size k be a positive integer, for example, k=26, meaning that the 26 closest points in 3D space surrounding the sampling point i are selected as the spatial neighborhood. For any sampling point j in the neighborhood, calculate two distance metrics: geometric distance and geometric distance. This is the straight-line distance between points i and j in the three-dimensional coordinate system. The material property distance is the three-dimensional material property vector corresponding to the two points. and The Euclidean distance between them. For example, if the material property vector of sampling point i... Let [0.8, 0.9, 0.3] represent the reinforcing steel, and let [0.8, 0.9, 0.3] be the material property vector of the nearest point j. Let [0.4, 0.5, 0.2] represent concrete; then the material properties differ significantly between them. The squared distance of the calculated material properties... and geometric distance squared Substitute the elements into the state transition element calculation formula. Preset weight coefficients. and This is used to balance the importance of material similarity and spatial proximity. If two points have similar material properties and are spatially close, the calculated... A value close to 1 indicates that the states are likely the same; conversely, if any difference is large, then... A value close to 0 indicates that the two points may belong to different materials or states, and there exists a state boundary. This is achieved by connecting each point i with all its neighboring points j. The values ​​are aggregated and normalized to form the state transition matrix of the entire model.

[0030] In an optional embodiment, the model's observation process matches different observation matrices for sampling points according to a preset material category, including: The materials are pre-categorized into concrete, reinforced steel, and voids; When the current sampling point is determined to belong to concrete based on the structural feature data, an observation matrix is ​​matched for that point. ; When the point is identified as a reinforcing bar, the observation matrix is ​​matched. ; When a hole is identified, the observation matrix is ​​matched for that point. ,in , , , , and These are the preset observation parameters corresponding to each material category.

[0031] Before updating the model's state, a preliminary material classification is performed for each sampling point. The classification is based on the structural feature data generated during the preprocessing stage, specifically the three-dimensional vectors of density, elastic modulus, and Poisson's ratio for each sampling point. This classification is achieved by setting threshold ranges; for example, concrete: density ∈ [0.6, 0.9], elastic modulus ∈ [0.5, 0.8], Poisson's ratio ∈ [0.2, 0.3]; steel reinforcement: density ∈ [0.85, 1.0], elastic modulus ∈ [0.8, 1.0], Poisson's ratio ∈ [0.3, 0.4]; voids: density ∈ [0, 0.1], elastic modulus ∈ [0, 0.1], Poisson's ratio ∈ [0, 0.1]. If a sampling point's attributes do not meet any of the category thresholds, it is temporarily classified as an unknown category, and the concrete observation matrix is ​​used. .

[0032] Once the material type of a sampling point is determined, a dedicated observation matrix is ​​assigned to that point. For example, if a sampling point is identified as concrete, subsequent observation update calculations will use the observation matrix. If another point is identified as a rebar, then use... The observation matrices are all diagonal matrices, with elements on the diagonal... These are observation parameters pre-set based on experimental data or experience, reflecting the uncertainty or noise level when observing hidden defects in that material. For example, observation parameters for concrete may correspond to higher uncertainty, while those for steel reinforcement may be lower. In this way, the model can incorporate prior knowledge of the material into the observation process, making the estimation of defects in different media more accurate, such as... Figure 2 .

[0033] In an optional embodiment, the state vector of the model is defined as a two-dimensional vector. ,in This represents the probability of a hidden defect existing at a sampling point. Health index represents the deviation of material properties from the corresponding category benchmark value; Health index The calculation method is as follows: Preset the baseline attribute vector for each material category, such as the baseline vector for healthy concrete. Healthy rebar reference vector Hollow with no baseline attribute Set the value to 0 and calculate the attribute vector of the sampling point. With the corresponding reference vector The relative Euclidean distance, i.e. 0.5 is a preset adjustment coefficient, which can be calibrated using measured data from standard material specimens. Specifically, it minimizes the mean square error between the predicted health index value and the actual health status. The larger the value, the worse the material health status.

[0034] S3, The hidden defect probability state value obtained above is used as a feature map and input into a defect detection network model using a segmented Swish activation function to obtain defect detection results; wherein, the segment threshold of the segmented Swish activation function is determined based on the statistical distribution of the feature map and the corresponding state estimation covariance; and, for input values ​​falling into different segment intervals divided by the segment threshold, a scaling factor corresponding to the width of the segment interval and the statistical characteristics of the input values ​​falling into the interval is used for calculation.

[0035] Specifically, the probability state values ​​of hidden defects at all spatial sampling points are organized into a two-dimensional or three-dimensional feature map. This feature map is input into a U-Net-like semantic segmentation network for defect detection. The activation functions of each convolutional layer in the network are replaced with segmented Swish activation functions. During the forward propagation of the network, statistical analysis is performed on the input feature map, and a histogram of the probability values ​​is calculated. Based on the histogram distribution, for example, the 80th and 95th percentiles are selected as the initial segmentation thresholds. and The input values ​​are divided into three intervals: low probability, medium probability, and high probability. Using the state estimation covariance map obtained above, the average covariance of the sampling points falling into each interval is calculated. If the average covariance of an interval is large, it indicates high uncertainty in the state estimation within that interval, and the boundary threshold is adjusted accordingly. After the threshold is determined, a scaling factor is calculated for each segmented interval. For the low-probability interval representing the background, a scaling factor less than 1 is set to suppress activation in that interval; for the uncertain medium-probability interval, a scaling factor greater than 1 is set to amplify the features of that interval and enhance the network's sensitivity to potential subtle defects; for the high-probability interval, a scaling factor close to 1 is set to preserve the original features of that interval, such as... Figure 3 The network learns and infers on the adjusted and activated feature maps, outputting pixel-level defect segmentation maps as defect detection results.

[0036] In an optional embodiment, the segmentation threshold of the segmented Swish activation function is determined based on the statistical distribution of the feature map and the corresponding state estimation covariance, including: Calculate the mean of the probability state values ​​of all hidden defects in the feature map. and standard deviation ; Calculate the average trace of the state estimation covariance matrix at the corresponding location in the feature map. ; And according to the formula and Set two segment thresholds and ,in This is the preset adjustment coefficient.

[0037] During the forward propagation of each layer in the neural network, before the feature map data is fed into the activation function, statistical analysis is performed on the feature map. This feature map represents the current network layer's probability state estimate of hidden defects at various spatial locations. The mean of all values ​​on this feature map is calculated. and standard deviation This reflects the overall distribution of activation values ​​in the current batch of data. It utilizes the state estimation covariance matrix associated with the state estimation algorithm. Each state estimate is accompanied by a covariance matrix, and the trace of this covariance matrix represents the magnitude of the uncertainty of the estimate. The traces of the covariance matrices at all locations are calculated, and the average is obtained. This value represents the overall confidence level of the current model in estimating the entire feature map. , , and a preset adjustment coefficient Substitute the values ​​into the threshold calculation formula. When the model has uncertainty... At higher levels, The term becomes smaller, making the threshold smaller. and To the mean When the model contracts, the linear region of the activation function narrows; conversely, when the model has high determinism, the linear region widens. This mechanism allows the shape of the activation function to be adjusted in real time according to the model's learning state and data characteristics.

[0038] In an optional embodiment, the step of calculating the scaling factor for input values ​​falling into different segment intervals divided by the segmentation threshold, using a scaling factor corresponding to the width of the segment interval and the statistical characteristics of the input values ​​falling into the interval, includes: For input value x, when When using a scaling factor ; when When using a scaling factor ; when When using a scaling factor ; And the input value x and the corresponding scaling factor Substitute into Swish activation function The calculation is performed in the following context, where the scaling factor is... Based on segmented intervals Width determined, scaling factor and The determination is based on the statistical characteristics of the input values ​​that fall into the corresponding intervals.

[0039] The threshold calculated using the above method and The domain of the input value is divided into three intervals: below... The low value area, between and The middle zone between, and above The high-value region. For any value x input to the activation function, determine the interval to which that value belongs.

[0040] Select a specific scaling factor based on the interval. For the input value in the middle area, the scaling factor Setting and interval width Correlation, such as inverse relationship, is used to control the gradient in that region. For low-value and high-value regions, the scaling factor... and The segmentation threshold is determined based on statistical data, such as variance, of all input values ​​falling within the two aforementioned intervals. For example, the segmentation threshold can be determined based on the feature map statistical distribution and the state estimation covariance. =0.2、 =0.7, corresponding to the interval width - =0.5; Scaling factor for the middle area The calculation is performed using a method inversely proportional to the interval width, with a proportionality coefficient k=1.0, i.e. =1.0 / 0.5=2.0, this setting enhances the gradient response in the intermediate region; subsequently, the variance of all input values ​​in the low-value region is calculated. The variance of all input values ​​in the high-value region is Low-value area scaling factor Set as High-value area scaling factor for Three sets of scaling factors were obtained. =2.0、 =10.0、 =14.3. The input value x and its corresponding scaling factor... Substitute them together into the expression of the Swish activation function In the middle, the activation calculation is completed. The piecewise scaling mechanism allows the activation function to have different response characteristics to inputs with different numerical ranges, enhancing the nonlinear expressive power of the model, such as... Figure 4 .

[0041] In an optional embodiment, the defect detection network model employing the piecewise Swish activation function includes: An encoder path consisting of N convolutional layers and M max pooling layers is used to extract features step by step; A decoder path consisting of N convolutional layers and M upsampling layers is used to progressively restore the feature map resolution. Skip connections are set between corresponding layers of the encoder and decoder paths to fuse shallow and deep features; Furthermore, all convolutional layers in the encoder and decoder paths employ the segmented Swish activation function, where N and M are preset positive integers.

[0042] The defect detection network model employs a symmetrical encoder-decoder structure. The input is a 3D feature map, for example, a single-channel feature map of 128×128×128×1, constructed from the hidden defect probability state values ​​of each spatial sampling point. The encoder path, responsible for feature extraction and spatial dimensionality reduction, consists of M downsampling modules cascaded together, for example, M=4. Each module contains several convolutional layers, such as two 3×3×3 convolutional layers followed by a 2×2×2 max-pooling layer. After passing through the encoder path, the spatial resolution of the feature map gradually decreases, for example, to 8×8×8, while the number of feature channels increases accordingly, thus detecting more abstract semantic information.

[0043] The decoder path is responsible for accurate defect localization and resolution restoration. It also consists of M upsampling modules, each connected to several convolutional layers via an upsampling layer (e.g., transposed convolution) that doubles the spatial size of the feature map. To address information loss during downsampling, the core of the U-Net structure is skip connections. These connections concatenate the feature maps output by each downsampling module in the encoder path to the output of the corresponding resolution upsampling module in the decoder path. This operation integrates the shallow, high-resolution detail features of the encoder with the deep, abstract semantic features of the decoder, improving the accuracy of the segmentation results. All convolutional layers in the network use the segmented Swish activation function described above, enhancing the network's ability to recognize complex defect patterns. The network output is a probability map of the same size as the input, where the value of each voxel represents the probability of a defect at that location. The training set for the U-Net network model consists of a large amount of 3D concrete structure simulation data and some non-destructive testing data with real defect annotations. Each training data point contains a 3D structural feature data volume as network input and a corresponding 3D defect label volume, which accurately marks the true location and shape of defects such as voids and cracks. The training process is end-to-end, with training data input into the network in batches. The network outputs a defect probability map, and the loss value between the network output and the true defect label volume is calculated using the Dice loss function or cross-entropy loss function. Subsequently, the Adam optimizer is used, with a learning rate of 1e-4 and weight decay of 1e-5. Based on this loss value, the weight parameters of all convolutional layers in the network are iteratively updated using the backpropagation algorithm until the model's performance metrics on independent validation sets, such as the intersection-over-union ratio (IoU), reach the preset convergence criterion.

[0044] In a second embodiment, the present invention also provides a construction concealed defect detection system, comprising the following modules: The generation module is used to acquire the raw data from the multi-channel sensors of the construction structure to be inspected, and to preprocess the data to generate structural feature data containing the material properties of each spatial sampling point. A construction module is used to construct a spatial state model, perform state estimation on the structural feature data, and obtain the hidden defect probability state value and the corresponding state estimation covariance of each spatial sampling point. The state transition process of the model determines the state transition matrix based on the material property gradient and geometric distance between the current sampling point and other sampling points in the spatial neighborhood. The observation process of the model matches different observation matrices for the sampling points according to a preset material category and updates the state in conjunction with the structural feature data. The calculation module is used to input the obtained hidden defect probability state value as a feature map into a defect detection network model using a piecewise Swish activation function to obtain defect detection results; wherein, the segment threshold of the piecewise Swish activation function is determined based on the statistical distribution of the feature map and the corresponding state estimation covariance; and, for input values ​​falling into different segment intervals divided by the segment threshold, a scaling factor corresponding to the width of the segment interval and the statistical characteristics of the input values ​​falling into the interval is used for calculation.

[0045] In this specification, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise limited, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the associated listed items.

[0046] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0047] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting hidden defects in construction, characterized in that, Includes the following steps: Acquire raw data from multi-channel sensors of the construction structure to be inspected, and preprocess it to generate structural feature data containing material properties of each spatial sampling point; A spatial state model is constructed, and the state of the structural feature data is estimated to obtain the hidden defect probability state value and the corresponding state estimation covariance of each spatial sampling point. The state transition process of the model determines the state transition matrix based on the material property gradient and geometric distance between the current sampling point and other sampling points in the spatial neighborhood. The observation process of the model matches different observation matrices for the sampling points according to the preset material category and updates the state in combination with the structural feature data. The hidden defect probability state values ​​obtained above are used as feature maps and input into a defect detection network model employing a segmented Swish activation function to obtain defect detection results. The segment threshold of the segmented Swish activation function is determined based on the statistical distribution of the feature map and the corresponding state estimation covariance. Furthermore, for input values ​​falling into different segment intervals defined by the segment threshold, a scaling factor corresponding to the width of the segment interval and the statistical characteristics of the input values ​​falling into the interval is used for calculation.

2. The method according to claim 1, characterized in that, The process involves acquiring raw data from multi-channel sensors of the construction structure to be inspected, preprocessing it, and generating structural feature data containing material properties of each spatial sampling point, including: A three-dimensional Gaussian filter is used to smooth and denoise the raw data from the multi-channel sensors. The denoised data is subjected to min-max normalization to linearly map all data values ​​to the interval [0,1]. Furthermore, the values ​​of density, elastic modulus, and Poisson's ratio are extracted from the normalized data for each spatial sampling point as material properties, which together constitute the structural feature data.

3. The method according to claim 1, characterized in that, The state transition process of the model determines the state transition matrix based on the material property gradient and geometric distance between the current sampling point and other sampling points in the spatial neighborhood, including: For the current sampling point i, select the k sampling points with the closest three-dimensional Euclidean distance to the point as the spatial neighborhood; Calculate the material property vector between the current sampling point i and any sampling point j in the neighborhood. and square Euclidean distance and the square of the geometric distance Before calculation, the three-dimensional geometric coordinates of the sampling points are subjected to min-max normalization to make... It becomes a dimensionless quantity; According to the formula Calculate the state transition elements, and all calculated elements. Together they form the state transition matrix, where k is a positive integer. and These are the preset weighting coefficients.

4. The method according to claim 1, characterized in that, The observation process of the model matches different observation matrices for sampling points according to a preset material category, including: The materials are pre-categorized into concrete, reinforced steel, and voids; When the current sampling point is determined to belong to concrete based on the structural feature data, an observation matrix is ​​matched for that point. ; When the point is identified as a reinforcing bar, the observation matrix is ​​matched. ; When a hole is identified, the observation matrix is ​​matched for that point. ,in , , , , and These are the preset observation parameters corresponding to each material category.

5. The method according to claim 1, characterized in that, The segmentation threshold of the segmented Swish activation function is determined based on the statistical distribution of the feature map and the corresponding state estimation covariance, including: Calculate the mean of the probability state values ​​of all hidden defects in the feature map. and standard deviation ; Calculate the average trace of the state estimation covariance matrix at the corresponding location in the feature map. ; And according to the formula and Set two segment thresholds and ,in This is the preset adjustment coefficient.

6. The method according to claim 5, characterized in that, The calculation of input values ​​falling into different segmented intervals defined by the segmentation threshold, using a scaling factor corresponding to the width of the segmented interval and the statistical characteristics of the input values ​​falling into the interval, includes: For input value x, when When using a scaling factor ; when When using a scaling factor ; When a scaling factor is employed and the input value x is multiplied by the corresponding scaling factor into the Swish activation function is calculated.

7. The method of claim 1, wherein, The defect detection network model employing the piecewise Swish activation function includes: An encoder path consisting of N convolutional layers and M max pooling layers is used to extract features step by step; A decoder path consisting of N convolutional layers and M upsampling layers is used to progressively restore the feature map resolution. Skip connections are set between corresponding layers in the encoder and decoder paths to fuse shallow and deep features; Furthermore, all convolutional layers in the encoder and decoder paths employ the segmented Swish activation function, where N and M are preset positive integers.

8. A construction concealment defect detection system characterized by, Includes the following modules: The generation module is used to acquire the raw data from the multi-channel sensors of the construction structure to be inspected, and to preprocess the data to generate structural feature data containing the material properties of each spatial sampling point. A construction module is used to construct a spatial state model, perform state estimation on the structural feature data, and obtain the hidden defect probability state value and the corresponding state estimation covariance of each spatial sampling point. The state transition process of the model determines the state transition matrix based on the material property gradient and geometric distance between the current sampling point and other sampling points in the spatial neighborhood. The observation process of the model matches different observation matrices for the sampling points according to a preset material category and updates the state in conjunction with the structural feature data. The calculation module is used to input the obtained hidden defect probability state value as a feature map into a defect detection network model using a piecewise Swish activation function to obtain defect detection results; wherein, the segment threshold of the piecewise Swish activation function is determined based on the statistical distribution of the feature map and the corresponding state estimation covariance; and, for input values ​​falling into different segment intervals divided by the segment threshold, a scaling factor corresponding to the width of the segment interval and the statistical characteristics of the input values ​​falling into the interval is used for calculation.

9. The system of claim 8, wherein, The process involves acquiring raw data from multi-channel sensors of the construction structure to be inspected, preprocessing it, and generating structural feature data containing material properties of each spatial sampling point, including: A three-dimensional Gaussian filter is used to smooth and denoise the raw data from the multi-channel sensors. The denoised data is subjected to min-max normalization to linearly map all data values ​​to the interval [0,1]. Furthermore, the values ​​of density, elastic modulus, and Poisson's ratio are extracted from the normalized data for each spatial sampling point as material properties, which together constitute the structural feature data.

10. The system according to claim 8, characterized in that, The state transition process of the model determines the state transition matrix based on the material property gradient and geometric distance between the current sampling point and other sampling points in the spatial neighborhood, including: For the current sampling point i, select the k sampling points with the closest three-dimensional Euclidean distance to the point as the spatial neighborhood; Calculate the material property vector between the current sampling point i and any sampling point j in the neighborhood. and square Euclidean distance and the square of the geometric distance Before calculation, the three-dimensional geometric coordinates of the sampling points are subjected to min-max normalization to make... It becomes a dimensionless quantity; According to the formula Calculate the state transition elements, and all calculated elements. Together they form the state transition matrix, where k is a positive integer. and These are the preset weighting coefficients.

Images