A deep learning-based rainwater pipe defect intelligent identification method

Abstract

The present application relates to the field of pipeline defect intelligent identification, and more particularly to a rainwater pipeline defect intelligent identification method based on deep learning. The method comprises the following steps: collecting multi-channel vibration signals of a flaw detection device and spatial coordinates of excitation points, denoising the signals, extracting relative propagation time delay, constructing a correlation matrix, taking coordinates as nodes, extracting energy features, converting the matrix into edge weights by using a kernel function to construct a spatial topology graph, inputting the topology graph into a graph convolution network model to output defect probability values of nodes, generating a discrete probability graph, extracting a continuous defect boundary contour by using a fitting function, mapping the contour to a three-dimensional twin model to output an estimated repair volume. The present application effectively reduces the interference of pipe wall sound wave dispersion, breaks through the limitations of complex topology correlation processing, realizes the shape boundary description of hidden defect areas, and provides quantitative data support for pipeline engineering repair.

Description

Technical Field

[0001] This invention relates to the field of intelligent identification of pipeline defects, and in particular to an intelligent identification method for rainwater pipeline defects based on deep learning. Background Technology

[0002] With increasing service life and the widespread adoption of trenchless repair technologies, urban stormwater pipes are prone to developing hidden structural defects such as voids or loose linings. Traditional visual inspection can only observe the pipe surface and cannot probe the true condition of the internal pipe wall. Therefore, vibration testing technology based on mechanical impact and acoustic sensing has become a key detection method. This technology excites elastic vibration waves to propagate in the composite pipe wall medium and uses a sensor array to collect multi-channel dynamic response signals. In recent years, the industry has gradually introduced deep learning technology, aiming to build neural network models to perform data mining and feature extraction on massive temporal vibration signals, in order to achieve automated and intelligent identification of hidden pipe defects.

[0003] However, composite pipe walls can cause strong acoustic dispersion, and environmental noise interference can easily cause signal distortion. Furthermore, existing methods are unable to handle the complex topological relationships of the non-Euclidean space of the pipe wall surface, making it difficult to effectively integrate the spatial acoustic propagation relationships between various detection points. This results in weak anti-interference capabilities of the system under complex operating conditions and low accuracy in identifying hidden defects. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a deep learning-based intelligent identification method for rainwater pipe defects, aiming to improve the existing methods' difficulty in handling the spatial topological correlation of pipe walls and their susceptibility to wave dispersion and noise interference, which leads to low identification accuracy.

[0005] This invention provides the following technical solution: a deep learning-based intelligent identification method for rainwater pipe defects, comprising:

[0006] S1. Control the flaw detection equipment to move in the rainwater pipe and continuously excite it, synchronously collect the multi-channel vibration signals of the peripheral sensor array in the flaw detection equipment, and record the spatial position coordinates of the excitation point corresponding to each excitation.

[0007] S2. The multi-channel vibration signal is denoised using a frequency domain decomposition algorithm, and the relative propagation delay between the excitation point and each sensor in the peripheral sensor array is extracted using a phase weighted cross-correlation algorithm to construct a time delay correlation matrix.

[0008] S3. Using the spatial location coordinates as spatial nodes, extract the energy characteristics of the spatial nodes based on the multi-channel vibration signal, and use the kernel function to transform the time delay correlation matrix into edge weights between adjacent spatial nodes to construct a spatial topology graph.

[0009] S4. Input the spatial topology graph into a pre-trained graph convolutional network model, aggregate the energy features and edge weights of adjacent spatial nodes, and output the defect probability value of each spatial node.

[0010] S5. Generate a discrete probability map based on the defect probability value of each spatial node, and use a closed curve fitting function to process the probability gradient features in the discrete probability map to extract the continuous defect boundary contour.

[0011] S6. Map the continuous defect boundary contour to a three-dimensional twin model, calculate the enclosing area of ​​the continuous defect boundary contour, and output the recognition result including the defect morphology and the estimated repair volume.

[0012] Preferably, in step S1, the step of synchronously acquiring the multi-channel vibration signals of the peripheral sensor array in the flaw detection equipment includes:

[0013] The central excitation device in the flaw detection equipment is controlled to apply mechanical excitation to the inner wall of the rainwater pipe, thereby generating elastic vibration waves that propagate in the medium of the pipe wall.

[0014] The dynamic response signal of the elastic vibration wave is captured using the peripheral sensor array and converted into an electrical signal sequence;

[0015] The electrical signal sequence is clock-synchronized and analog-to-digital converted to generate the multi-channel vibration signal containing timestamps and amplitude data.

[0016] Preferably, in step S2, the step of extracting the relative propagation delay between the excitation point and each sensor in the peripheral sensor array includes:

[0017] The noise-reduced multi-channel vibration signal is converted to the frequency domain, and the cross-power spectral density between the excitation source signal corresponding to the excitation point and the received signal corresponding to each sensor is calculated.

[0018] A phase weighting function is introduced to perform frequency domain whitening on the cross-power spectral density in order to suppress dispersion interference caused by signal propagation;

[0019] Perform an inverse time-domain transform on the whitened spectral data and extract the time index of the extreme point in the cross-correlation function as the relative propagation delay.

[0020] Preferably, in step S3, the step of extracting the energy features of the spatial node includes:

[0021] Based on the triggering time of the excitation point, the multi-channel vibration signal is extracted by a preset time window to obtain a single excitation signal sequence.

[0022] Within the single excitation signal sequence, the time-domain amplitude integral is extracted as the time-domain energy feature, and the main frequency band component of the spectrum is extracted as the frequency domain distribution feature.

[0023] The energy features of the spatial node are generated by splicing and normalizing the time-domain energy features and the frequency-domain distribution features.

[0024] Preferably, in step S3, the step of using a kernel function to transform the time-delay correlation matrix into edge weights between adjacent spatial nodes includes:

[0025] Based on the spatial location coordinates of each of the spatial nodes, calculate the spatial physical distance between adjacent spatial nodes;

[0026] The kernel function is used to perform a nonlinear mapping on the ratio of the time delay value in the time delay correlation matrix to the corresponding spatial physical distance;

[0027] A preset truncation threshold is introduced to filter out associated edges in the mapping result that are less than the preset truncation threshold, and the edge weights that characterize the topological connection strength between nodes are output.

[0028] Preferably, in step S4, the step of aggregating the energy characteristics and edge weights of adjacent spatial nodes includes:

[0029] A Laplace matrix is ​​constructed based on the edge weights between adjacent spatial nodes, and the Laplace matrix is ​​then normalized.

[0030] In the graph convolutional network model, the normalized Laplacian matrix is ​​used to transfer features between adjacent spatial nodes to aggregate the energy features;

[0031] The aggregated features are mapped to a preset numerical range using an activation function, and the defect probability value of each spatial node is output.

[0032] Preferably, in step S5, the step of generating the discrete probability map includes:

[0033] Construct a two-dimensional cylindrical expansion coordinate system and establish a discrete mesh matrix describing the inner wall of the rainwater pipe;

[0034] Based on the coordinate mapping relationship, the spatial position coordinates of each spatial node and the defect probability value are projected onto the corresponding grid points of the discrete grid matrix;

[0035] Using a spatial interpolation algorithm, smooth transition interpolation is performed on unprojected grid points based on grid points with projected defect probability values ​​to generate the discrete probability map containing the global probability distribution.

[0036] Preferably, in step S5, the step of processing the probability gradient features in the discrete probability map includes:

[0037] Calculate the differential derivative of the discrete probability map and extract the extreme points that characterize the probability gradient features of the defect interface;

[0038] Using the envelope of the extreme point as the initial contour, an active contour functional containing an internal tension constraint term and an external gradient attraction term is constructed as the fitting function for the closed curve.

[0039] The continuous defect boundary profile is extracted by minimizing the computation to drive the active profile functional to shrink and close towards the extreme value region.

[0040] Preferably, in step S6, the step of calculating the enclosing area of ​​the continuous defect boundary profile includes:

[0041] Obtain the pipe geometry parameters in the three-dimensional twin model, and inversely map the continuous defect boundary contour to a three-dimensional spatial surface based on the pipe geometry parameters;

[0042] The surface integral algorithm is used to perform surface integral along the mapped continuous defect boundary contour on the three-dimensional space surface to calculate the enclosed area;

[0043] The estimated repair volume, including the defect morphology, is calculated by multiplying the enclosed area with a preset pipe wall depth parameter.

[0044] The present invention has the following beneficial effects:

[0045] 1. In this invention, the relative propagation delay of the vibration signal is transformed into a spatial topology map and feature aggregation is performed using a graph convolutional network, which effectively reduces the dispersion interference of the pipe wall sound wave. This enables the system to accurately predict hidden defects in harsh noise environments by relying on the node message passing mechanism, thereby improving the stability of intelligent identification.

[0046] 2. In this invention, a discrete probability map is generated by spatial interpolation and fitted by an active contour functional. This transforms the scattered predicted probability distribution into a closed and geometrically smooth continuous curve, which improves the problem that traditional mechanical flaw detection is difficult to accurately depict the physical boundary of the defect. This enables high-precision morphological boundary characterization of the ruptured area of ​​the pipe wall.

[0047] 3. In this invention, the continuous defect boundary contour is reverse-mapped to a three-dimensional twin surface and the surface integral algorithm is used to calculate the actual enclosed area. Combined with the depth parameter, the estimated repair volume is directly calculated. The abstract underlying reasoning results are transformed into engineering surveying data with clear physical dimensions, providing key material calculation basis for trenchless grouting repair construction of municipal pipelines. Attached Figure Description

[0048] Figure 1 This is a flowchart of a deep learning-based intelligent identification method for rainwater pipe defects proposed in this invention.

[0049] Figure 2 This is a flowchart of the relative propagation delay extraction based on phase weighting proposed in this invention;

[0050] Figure 3 This is a flowchart illustrating the construction process of the non-Euclidean geometric space topology graph proposed in this invention.

[0051] Figure 4 This is a flowchart illustrating the feature aggregation and deduction process of the graph convolutional network proposed in this invention.

[0052] Figure 5 This is a flowchart of the continuous boundary extraction process proposed in this invention using the active contour functional.

[0053] Figure 6 This is a flowchart of the twin surface mapping and volume calculation proposed in this invention. Detailed Implementation

[0054] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] In embodiments of the present invention, the present invention provides a method for intelligent identification of rainwater pipe defects based on deep learning, such as... Figure 1 As shown, it includes:

[0056] S1. Control the flaw detection equipment to move in the rainwater pipe and continuously excite it, synchronously collect the multi-channel vibration signals of the peripheral sensor array in the flaw detection equipment, and record the spatial position coordinates of the excitation point corresponding to each excitation.

[0057] Furthermore, step S1, the step of synchronously acquiring multi-channel vibration signals from the peripheral sensor array in the flaw detection equipment, includes:

[0058] The central excitation device in the flaw detection equipment is controlled to apply mechanical excitation to the inner wall of the rainwater pipe, thereby generating elastic vibration waves that propagate in the medium of the pipe wall.

[0059] The dynamic response signal of elastic vibration wave is captured by an external sensor array and converted into an electrical signal sequence;

[0060] The electrical signal sequence is clock-synchronized and analog-to-digital converted to generate a multi-channel vibration signal containing timestamps and amplitude data.

[0061] Specifically, the flaw detection equipment travels within the rainwater pipe to be inspected at a set speed. During this process, the central excitation device mounted on the flaw detection equipment uses an electromagnetic impact device to apply transient mechanical excitation to the inner wall of the rainwater pipe. The impact force of the mechanical excitation acts directly on the pipe wall, generating elastic vibration waves that spread outwards in the pipe wall medium and the surrounding external structure.

[0062] The peripheral sensor array, arranged in a geometric array around the central excitation device, employs high-frequency piezoelectric accelerometers. When the elastic vibration wave propagates along the pipe wall medium to the location of each sensor, the peripheral sensor array captures the dynamic response signal of the elastic vibration wave changing over time, and uses piezoelectric conversion elements to convert the captured physical and mechanical vibration into a continuous analog electrical signal sequence in real time.

[0063] After receiving the electrical signal sequence, the multi-channel high-precision data acquisition board inside the flaw detection equipment uses a global hardware triggering mechanism to perform clock synchronization and analog-to-digital conversion on the signal. At the moment the central excitation device contacts the pipe wall and triggers an impact, the system records the absolute position of the current impact point within the rainwater pipe using an odometer and inertial measurement unit, using this as the spatial coordinates of the excitation point. Simultaneously, the system discretizes the analog signals from each channel according to the set sampling frequency, generating a multi-channel vibration signal containing timestamps and digital amplitude data.

[0064] Let the total number of sensors in the peripheral sensor array be... The sampling frequency set for the analog-to-digital conversion process is... In this scheme, transient capture of high-frequency elastic vibration waves is addressed. Preferred Then the discrete sampling time step is Within a single mechanical excitation cycle, in this scheme, the optimal effective time window for single excitation data acquisition is [missing information]. The corresponding total number of sampling points Preferred The generated multi-channel vibration signal, containing timestamps and amplitude data, is represented as a discrete-time series matrix. :

[0065] ;

[0066] In the matrix above, the discrete timestamp is defined as Column vectors Indicates a specific timestamp The synchronous sampling data matrix of the entire peripheral sensor array at any given time:

[0067] ;

[0068] In the formula, A unified start timestamp triggered by the physical striking action of the central excitation device. This refers to the total number of sampling points within a single excitation acquisition time window. Indicates the first A peripheral sensor at a timestamp The digital amplitude values ​​of the elastic vibration waves of the rainwater pipes are collected and converted at all times. This represents the vector transpose operation. The spatial position coordinates corresponding to each excitation are represented as a three-dimensional vector. , respectively represent the axial travel distance, horizontal offset, and longitudinal elevation position of the excitation point in the physical space of the rainwater pipe.

[0069] The implementation of this step provides the pipeline defect identification system with basic data on the acoustic response of the pipe wall, which has a unified time reference and precise spatial location mapping.

[0070] S2. The frequency domain decomposition algorithm is used to denoise the multi-channel vibration signal, and the phase weighted cross-correlation algorithm is used to extract the relative propagation delay between the excitation point and each sensor in the peripheral sensor array in order to construct the time delay correlation matrix.

[0071] Further, step S2, the step of extracting the relative propagation delay between the excitation point and each sensor in the peripheral sensor array, includes:

[0072] The noise-reduced multi-channel vibration signal is converted to the frequency domain, and the cross power spectral density between the excitation source signal corresponding to the excitation point and the received signal corresponding to each sensor is calculated.

[0073] A phase weighting function is introduced to perform frequency domain whitening on the cross-power spectral density in order to suppress dispersion interference caused by signal propagation;

[0074] Perform an inverse time-domain transform on the whitened spectral data and extract the time index of the extreme point in the cross-correlation function as the relative propagation delay.

[0075] Specifically, after acquiring the multi-channel vibration signals, the system employs a frequency domain decomposition algorithm for noise reduction. In practice, the time-domain multi-channel vibration signals are converted to the frequency domain using discrete Fourier transform. A bandpass filter range is set based on the preset effective acoustic frequency band of the rainwater pipe medium. In this scheme, for the concrete and resin composite pipe wall structure, the lower limit frequency of the bandpass filter is preferably... The upper limit frequency is preferably The low-frequency water flow interference noise and high-frequency equipment electromagnetic noise components are filtered out, and then restored to the time-domain signal through inverse Fourier transform. The excitation source signal after noise reduction is defined as follows: After noise reduction The received signals corresponding to each peripheral sensor are .

[0076] The noise-reduced excitation source signal and the corresponding received signals from each sensor are transformed into the frequency domain using Fourier transform, respectively, to obtain the corresponding frequency domain signals. and Based on the frequency domain signal generated above, the cross-power spectral density between the excitation source signal corresponding to the excitation point and the received signals corresponding to each sensor is calculated. :

[0077] ;

[0078] In the formula, Indicates the angular frequency of the signal. Indicates the first The frequency domain signals corresponding to the signals received by the peripheral sensors are complex conjugates. This cross-power spectral density is directly related to the phase difference information of the elastic wave propagating in the rainwater pipe medium.

[0079] Because the composite wall structure of rainwater pipes causes dispersion effects during the propagation of elastic vibration waves, the system introduces a phase weighting function to perform frequency domain whitening of the cross-power spectral density. Using a phase transform weighting rule, the reciprocal of the amplitude of the cross-power spectral density is assigned as a weighting parameter to calculate the whitened cross-spectral data. :

[0080] ;

[0081] In the formula, This represents the absolute value of the cross-power spectral density amplitude. Through the above frequency domain whitening operation, the system forcibly balances the amplitude weights of all frequency components, effectively suppressing dispersion interference caused by uneven attenuation of the pipe wall material, and retaining only the pure phase shift information representing the true arrival time.

[0082] Perform inverse time-domain transform on the whitened spectral data to extract the generalized cross-correlation function in the corresponding time domain. :

[0083] ;

[0084] In the formula, Indicates the propagation time offset. The imaginary unit is used. The system globally searches the generalized cross-correlation function sequence within the time domain for the time index position corresponding to the maximum extreme point, and uses this as the excitation point up to the [number missing]. Relative propagation delay of the peripheral sensors :

[0085] ;

[0086] The system continuously traverses all independent sensor nodes in the peripheral sensor array, extracts the relative propagation delay features corresponding to each data channel, and performs column vector combination calculations according to the physical arrangement sequence of the sensors in the array to construct a delay correlation matrix containing the relative time difference features of each detection azimuth. :

[0087] ;

[0088] In the formula, This represents the total physical number of sensors within the peripheral sensor array. This is the transpose symbol for a matrix.

[0089] This step eliminates waveform distortion interference caused by complex pipeline environment noise and medium dispersion, and accurately extracts the spatial acoustic propagation time delay characteristics that reflect the internal impedance changes of the pipe wall structure.

[0090] S3. Using spatial location coordinates as spatial nodes, extract the energy characteristics of spatial nodes based on multi-channel vibration signals, and use kernel functions to transform the time delay correlation matrix into edge weights between adjacent spatial nodes to construct a spatial topology graph.

[0091] Furthermore, step S3, the step of extracting the energy features of spatial nodes, includes:

[0092] Based on the trigger time of the excitation point, the multi-channel vibration signal is extracted by a preset time window to obtain a single excitation signal sequence;

[0093] Within a single excitation signal sequence, the time-domain amplitude integral is extracted as the time-domain energy feature, and the main frequency band component of the spectrum is extracted as the frequency-domain distribution feature.

[0094] The energy characteristics of spatial nodes are generated by splicing and normalizing the time-domain energy characteristics and the frequency-domain distribution characteristics.

[0095] Furthermore, step S3, which involves using a kernel function to transform the time-delay correlation matrix into edge weights between adjacent spatial nodes, includes:

[0096] Calculate the physical distance between adjacent spatial nodes based on the spatial coordinates of each spatial node.

[0097] A kernel function is used to perform a nonlinear mapping on the ratio of the time delay value in the time delay correlation matrix to the corresponding spatial physical distance;

[0098] Introducing a preset truncation threshold, filtering out associated edges in the mapping result that are less than the preset truncation threshold, and outputting edge weights that characterize the strength of topological connections between nodes.

[0099] Specifically, the system defines the spatial coordinates corresponding to each recorded excitation as spatial nodes in a spatial network. Using the trigger time of the excitation point recorded by the hardware trigger as the starting point of the time reference, a preset time window length is set based on the typical attenuation period of sound waves within the rainwater pipe wall material. The system performs a segmentation operation on the multi-channel vibration signals to obtain a sequence of single excitation signals located within this preset time window.

[0100] Within a single excitation signal sequence, the system extracts the time-domain amplitude integral as the time-domain energy feature by calculating the sum of squares of the amplitudes at discrete sampling points of the signal. :

[0101] ;

[0102] In the formula, The first in a single excitation signal sequence Amplitude at discrete sampling points This represents the total number of sampling points within a preset time window. Simultaneously, the system performs a Discrete Fourier Transform on the single excitation signal sequence to obtain its spectral distribution. Within the dominant frequency band characterizing the pipe structure response, the spectral energy is summed, and the dominant frequency band component is extracted as the frequency domain distribution feature. :

[0103] ;

[0104] In the formula, For frequency sequence indexing, and These are the lower and upper limits of the main frequency band, set based on the resonance characteristics of the rainwater pipe structure. In this scheme, to accurately extract the resonance characteristics of pipe wall voiding, Preferred , Preferred , This represents the amplitude spectrum value at the corresponding frequency.

[0105] The system concatenates the extracted time-domain energy features and frequency-domain distribution features into column vectors to form a joint feature parameter. The joint feature parameter is then linearly normalized using a maximum-minimum scaling algorithm, mapping its value to a fixed interval between zero and one, thereby generating an energy feature that characterizes the acoustically dense state of the pipe wall at the spatial node.

[0106] For any two adjacent spatial nodes in the graph construction process With spatial nodes The system calculates the physical distance between adjacent spatial nodes based on their included spatial coordinates. :

[0107] ;

[0108] In the formula, and They are spatial nodes With spatial nodes The location coordinate components of the rainwater pipe in the physical space. The system extracts the time delay values ​​between corresponding nodes in the time delay correlation matrix. A Gaussian radial basis function is used to perform a nonlinear mapping on the ratio of the time delay value to the corresponding spatial physical distance, and the initial edge weights are calculated. :

[0109] ;

[0110] In the formula, To control the standard deviation parameter of the kernel function mapping distribution width, this scheme balances the sparsity and connectivity of the network. Preferred The ratio of time delay to physical distance The acoustic slowness characterizes the local propagation of elastic waves in the rainwater pipe medium.

[0111] The system introduces a preset cutoff threshold. In this plan, Preferred The initial edge weights are conditionally determined, and associated edges in the mapping result that are less than a preset truncation threshold are filtered out to reduce network complexity. When the value is less than the preset truncation threshold, the network connection is blocked and the value is set to zero. For initial edge weights that are greater than or equal to the preset truncation threshold, their mapping values ​​are retained, and the edge weights that characterize the spatial acoustic topological connection strength between rainwater pipe network nodes are output to construct a spatial topology graph.

[0112] This step converts discrete pipe wall vibration data into mathematical graph structural features, quantifying the acoustic connectivity relationship between various physical detection points in the pipe.

[0113] S4. Input the spatial topology graph into the pre-trained graph convolutional network model, aggregate the energy features and edge weights of adjacent spatial nodes, and output the defect probability value of each spatial node.

[0114] Furthermore, step S4, which involves aggregating the energy characteristics and edge weights of adjacent spatial nodes, includes:

[0115] Construct a Laplacian matrix based on the edge weights between adjacent spatial nodes, and then normalize the Laplacian matrix.

[0116] In graph convolutional network models, the normalized Laplacian matrix is ​​used to transfer features between adjacent spatial nodes to aggregate energy features;

[0117] The aggregated features are mapped to a preset numerical range using an activation function, and the defect probability value of each spatial node is output.

[0118] Specifically, the system constructs an adjacency matrix representing the acoustic topological relationships of the pipe wall based on the edge weights between adjacent spatial nodes. To preserve the tube wall detection energy characteristics of the spatial nodes during feature transfer, the system will use the identity matrix... Superimposed on the adjacency matrix to form an augmented adjacency matrix that introduces self-loops. Then, the diagonal matrix corresponding to this augmented adjacency matrix is ​​calculated. The diagonal elements of the degree matrix are equal to the sum of the weights of the corresponding row nodes in the augmented adjacency matrix. Based on the above data, a matrix is ​​constructed, and the system performs a symmetric normalization operation on it to calculate the normalized Laplacian matrix. :

[0119] ;

[0120] In the formula, the Laplace matrix The internal elements rigorously quantify the standardized acoustic feature transmission probability between any two physically connected nodes on the wall of the rainwater pipe from a mathematical perspective.

[0121] The system inputs spatial topology graph data, including the spatial locations of pipe walls and edge weights, into a pre-trained graph convolutional network model. The energy features of each spatial node extracted earlier are combined according to the node sequence to form an initial feature matrix. The normalized Laplacian matrix is ​​used to perform feature transfer and weighted summation between connected adjacent spatial nodes. (Figure: The convolutional network model...) The feature aggregation calculation process of a layer is represented as follows:

[0122] ;

[0123] In the formula, and These represent the spatial nodes at the th, respectively. Layer input and the first The implicit feature matrix output by the layer. This represents the learnable network weight parameter matrix obtained through pre-training in the current hidden layer of the graph convolutional network model. This is a linear rectification activation operation. During this propagation process, the independent acoustic characteristics of the local pipe wall detection point and the structural response characteristics of the surrounding physical neighboring nodes complete a deep spatial dimension aggregation.

[0124] After multi-layer feature aggregation operations in the graph network, the final output of the graph convolutional network model is reduced to a single-channel final node feature sequence. The system uses a logistic regression activation function to perform a non-linear mapping operation on the final node feature sequence, forcibly constraining the unbounded network feature scalars to a preset numerical range of zero to one. For the _th ... Each spatial node has a defect probability value. The mapping formula is:

[0125] ;

[0126] In the formula, Representing the The final network feature scalar value output after the spatial nodes are aggregated through the graph network, and the exponential operation term. By incorporating a constant term distribution, features are mapped to probability values. The final output is the defect probability value. The statistical confidence level of the actual physical area of ​​the rainwater pipe at the corresponding spatial coordinates is directly characterized by the occurrence of voids or rupture defects.

[0127] This step utilizes the message-passing property of graph convolutional networks to fuse the associated physical responses of the local pipe wall and its surroundings, thus achieving accurate quantitative probability prediction of hidden defects.

[0128] S5. Generate a discrete probability map based on the defect probability values ​​of each spatial node, and use a closed curve fitting function to process the probability gradient features in the discrete probability map to extract the continuous defect boundary contour.

[0129] Further, step S5, the step of generating the discrete probability graph, includes:

[0130] Construct a two-dimensional cylindrical expansion coordinate system and establish a discrete mesh matrix describing the inner wall of the rainwater pipe;

[0131] Based on the coordinate mapping relationship, the spatial location coordinates and defect probability values ​​of each spatial node are projected onto the corresponding grid points of the discrete grid matrix;

[0132] Using a spatial interpolation algorithm, unprojected grid points are smoothly interpolated based on grid points with projected defect probability values ​​to generate a discrete probability map containing the global probability distribution.

[0133] Furthermore, step S5, the step of processing the probability gradient features in the discrete probability graph, includes:

[0134] Calculate the difference derivative of the discrete probability map and extract the extreme points that characterize the probability gradient features of the defect interface.

[0135] Using the envelope where the extreme point is located as the initial contour, we construct an active contour functional containing internal tension constraint terms and external gradient attraction terms as a closed curve fitting function.

[0136] By minimizing the computationally driven active profile functional to shrink and close towards the extreme region, the continuous defect boundary profile is extracted.

[0137] Specifically, the design inner diameter parameters of the rainwater pipe to be tested are obtained. The three-dimensional cylindrical surface of the pipe wall is cut along the longitudinal geometric center axis of the pipe and unfolded into a two-dimensional plane, thereby constructing a two-dimensional cylindrical unfolded coordinate system. The horizontal axis component of this coordinate system is defined. The axial travel distance of the pipeline, with the longitudinal component as the longitudinal axis. Let be the circumferential arc length of the pipe wall. A discrete mesh matrix describing the inner wall of the rainwater pipe is established according to the spatial sampling resolution set by the project. Based on the coordinate mapping relationship from cylinder to plane, the spatial coordinates of each spatial node in three-dimensional physical space are converted into the horizontal and vertical index positions of the discrete mesh matrix, and the defect probability value corresponding to each spatial node is directly assigned to the corresponding mesh point in the discrete mesh matrix.

[0138] For blank grid points in the discrete grid matrix that are not projected, the system employs an inverse distance weighted spatial interpolation algorithm for smooth transition interpolation calculations. The interpolation defect probability value at unprojected grid points in the two-dimensional coordinate plane is defined as... The interpolation calculation formula is as follows:

[0139] ;

[0140] In the formula, This represents the total number of grid points containing the projected defect probability values ​​within a set local search radius. In this scheme, the local search radius is preferably... , Preferably, it is a neighboring area within this radius. A known grid point, For the first The known defect probability value for each projected grid point. For the unprojected grid point to the th The system calculates the Euclidean distance of each projected grid point in the expanded coordinate system within a two-dimensional plane. It then iterates through all unprojected grid points, performing the above calculation to generate a discrete probability map covering the entire pipe wall surface with continuous data. .

[0141] The system performs discrete difference derivative calculations on the generated discrete probability map. First-order partial derivatives along the axial and circumferential directions of the pipe wall are calculated using the image spatial partial derivative operator, and the magnitude field of the extracted two-dimensional probability gradient features is then fused. :

[0142] ;

[0143] In the formula, the partial derivative components characterize the degree of abrupt change in probability values ​​between the normal pipe wall region and the suspected void / rupture region of the rainwater pipe. The system extracts the set of grid coordinates where the local amplitude reaches its maximum in the two-dimensional amplitude field, and uses these as the extreme points characterizing the probability gradient features of the defect interface.

[0144] The geometric envelope formed by connecting adjacent extreme points in space is used as the initial contour. A parameterized active contour functional is constructed as the closed curve fitting function to refine the defect boundary extraction. The parameterized equation for the two-dimensional closed curve is defined as follows: ,variable The normalized arc length parameter is used along the curve direction. The total energy function of the constructed active contour functional is... as follows:

[0145] ;

[0146] In the formula, the combination of the squared terms of the first-order partial derivatives and the squared terms of the second-order partial derivatives constitutes the internal tension constraint term, and the squared term of the gradient magnitude led by the negative sign is the external gradient attraction term. and These are the penalty weighting coefficients for controlling the elastic contraction tension and bending stiffness at the defect boundary, respectively. In this scheme, Preferred , Preferred , To control the weighting parameters of the external probability gradient field's attraction intensity, this scheme ensures that the boundary converges towards the high gradient region at the defect edge. Preferred The system iteratively solves for the minimum value of the functional using the variational method, driving the closed curve fitting function to continuously shrink towards the extreme value region of the probability gradient and close at both ends, thus extracting a geometrically smooth and coherent continuous defect boundary profile.

[0147] This step transforms the scattered probability thermal distribution output by the graph network into mathematically closed geometric lines, accurately outlining the actual two-dimensional physical contours of hidden defects in rainwater pipes.

[0148] S6. Map the continuous defect boundary contour to the three-dimensional twin model, calculate the enclosing area of ​​the continuous defect boundary contour, and output the recognition result containing the defect morphology and the estimated repair volume.

[0149] Further, step S6, the step of calculating the enclosing area of ​​the continuous defect boundary profile, includes:

[0150] Obtain the pipe geometry parameters in the 3D twin model, and inversely map the continuous defect boundary contour to the 3D spatial surface based on the pipe geometry parameters;

[0151] The surface integration algorithm is used to perform surface integration on a three-dimensional spatial surface along the mapped continuous defect boundary contour to calculate the enclosed area.

[0152] The estimated repair volume, including the defect morphology, is calculated by multiplying the enclosed area with the preset pipe wall depth parameter.

[0153] Specifically, the system extracts the basic pipe geometric parameters from the three-dimensional twin model of the rainwater pipes, mainly obtaining the design inner diameter radius of the rainwater pipes. In this plan, we take the standard branch pipe for trenchless repair in municipal engineering as an example. Preferred Based on the pipeline's geometric parameters, the system extracts the continuous defect boundary profile in a two-dimensional plane unfolded coordinate system. Inversely map the image onto the real 3D space. Set the contour mapping vector on the 3D surface as... Based on the physical mapping relationship between the two-dimensional vertical axis corresponding to the axial distance of the pipe and the two-dimensional horizontal axis corresponding to the circumferential arc length of the pipe wall, the system wraps it onto a three-dimensional spatial surface. The geometric transformation process of the inverse mapping is as follows:

[0154] ;

[0155] In the formula, Represents the physical inner radius of the rainwater pipe, a variable. This is the physical central angle corresponding to the profile arc length parameter on the circular cross-section of the pipe.

[0156] After mapping and reconstructing the continuous defect boundary contour on a three-dimensional curved surface, the system uses a surface integral algorithm to calculate the actual physical area enclosed by the closed boundary. Since the three-dimensional cylindrical surface is a Euclidean developable surface with zero Gaussian curvature, the surface integral on the cylindrical surface is numerically equivalent to the line integral of a closed curve on a two-dimensional plane. The system applies Green's theorem to perform surface integral operations along the mapped continuous defect boundary contour to calculate the actual enclosed area. :

[0157] ;

[0158] In the formula, the closed integration domain Differential operator representing the continuous defect boundary profile with its ends connected. and These represent the first derivatives of the coordinate components of the boundary profile along the parameterized direction. The enclosed area calculated by the above integral summation accurately reflects the physical surface area of ​​the stormwater pipe wall where structural damage has occurred.

[0159] The system reads the preset pipe wall depth parameters. In this plan, a typical void repair assessment is conducted between the resin liner and the old pipe wall. Preferred This data originates from the baseline of the old pipe wall thickness established in the project or the calculated value of the local void depth detected by external radar equipment. The system performs a direct multiplication operation between the enclosed area obtained by surface integration and the pipe wall depth parameter to calculate the estimated repair volume in three-dimensional space. :

[0160] ;

[0161] In the formula, the estimated repair volume is... This represents the theoretical material filling volume required to fill the structural defect in the pipe wall during the project. The system integrates and packages the reconstructed 3D defect topology data with the estimated repair volume data, and outputs it as an intelligent recognition result to the terminal interactive interface of the 3D twin model for display.

[0162] This step converts the topological profile generated by the algorithm into physical geometric parameters with engineering dimensions, outputting accurate quantitative indicators of area and volume, providing direct data support for material accounting for municipal pipeline network repair.

[0163] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent identification of rainwater pipe defects based on deep learning, characterized in that, include: S1. Control the flaw detection equipment to move in the rainwater pipe and continuously excite it, synchronously collect the multi-channel vibration signals of the peripheral sensor array in the flaw detection equipment, and record the spatial position coordinates of the excitation point corresponding to each excitation. S2. The multi-channel vibration signal is denoised using a frequency domain decomposition algorithm, and the relative propagation delay between the excitation point and each sensor in the peripheral sensor array is extracted using a phase weighted cross-correlation algorithm to construct a time delay correlation matrix. S3. Using the spatial location coordinates as spatial nodes, extract the energy characteristics of the spatial nodes based on the multi-channel vibration signal, and use the kernel function to transform the time delay correlation matrix into edge weights between adjacent spatial nodes to construct a spatial topology graph. S4. Input the spatial topology graph into a pre-trained graph convolutional network model, aggregate the energy features and edge weights of adjacent spatial nodes, and output the defect probability value of each spatial node. S5. Generate a discrete probability map based on the defect probability value of each spatial node, and use a closed curve fitting function to process the probability gradient features in the discrete probability map to extract the continuous defect boundary contour. S6. Map the continuous defect boundary contour to a three-dimensional twin model, calculate the enclosing area of ​​the continuous defect boundary contour, and output the recognition result including the defect morphology and the estimated repair volume.

2. The intelligent identification method for rainwater pipe defects based on deep learning according to claim 1, characterized in that, Step S1, the step of synchronously acquiring multi-channel vibration signals from the peripheral sensor array in the flaw detection equipment, includes: The central excitation device in the flaw detection equipment is controlled to apply mechanical excitation to the inner wall of the rainwater pipe, thereby generating elastic vibration waves that propagate in the medium of the pipe wall. The dynamic response signal of the elastic vibration wave is captured using the peripheral sensor array and converted into an electrical signal sequence; The electrical signal sequence is clock-synchronized and analog-to-digital converted to generate the multi-channel vibration signal containing timestamps and amplitude data.

3. The intelligent identification method for rainwater pipe defects based on deep learning according to claim 1, characterized in that, Step S2, the step of extracting the relative propagation delay between the excitation point and each sensor in the peripheral sensor array, includes: The noise-reduced multi-channel vibration signal is converted to the frequency domain, and the cross-power spectral density between the excitation source signal corresponding to the excitation point and the received signal corresponding to each sensor is calculated. A phase weighting function is introduced to perform frequency domain whitening on the cross-power spectral density in order to suppress dispersion interference caused by signal propagation; Perform an inverse time-domain transform on the whitened spectral data and extract the time index of the extreme point in the cross-correlation function as the relative propagation delay.

4. The intelligent identification method for rainwater pipe defects based on deep learning according to claim 1, characterized in that, Step S3, the step of extracting the energy features of the spatial nodes, includes: Based on the triggering time of the excitation point, the multi-channel vibration signal is extracted by a preset time window to obtain a single excitation signal sequence. Within the single excitation signal sequence, the time-domain amplitude integral is extracted as the time-domain energy feature, and the main frequency band component of the spectrum is extracted as the frequency domain distribution feature. The energy features of the spatial node are generated by splicing and normalizing the time-domain energy features and the frequency-domain distribution features.

5. The intelligent identification method for rainwater pipe defects based on deep learning according to claim 1, characterized in that, Step S3, the step of transforming the time delay correlation matrix into edge weights between adjacent spatial nodes using a kernel function, includes: Based on the spatial location coordinates of each of the spatial nodes, calculate the spatial physical distance between adjacent spatial nodes; The kernel function is used to perform a nonlinear mapping on the ratio of the time delay value in the time delay correlation matrix to the corresponding spatial physical distance; A preset truncation threshold is introduced to filter out associated edges in the mapping result that are less than the preset truncation threshold, and the edge weights that characterize the topological connection strength between nodes are output.

6. The intelligent identification method for rainwater pipe defects based on deep learning according to claim 1, characterized in that, In step S4, the step of aggregating the energy features and edge weights of adjacent spatial nodes includes: A Laplace matrix is ​​constructed based on the edge weights between adjacent spatial nodes, and the Laplace matrix is ​​then normalized. In the graph convolutional network model, the normalized Laplacian matrix is ​​used to transfer features between adjacent spatial nodes to aggregate the energy features; The aggregated features are mapped to a preset numerical range using an activation function, and the defect probability value of each spatial node is output.

7. The intelligent identification method for rainwater pipe defects based on deep learning according to claim 1, characterized in that, Step S5, the step of generating the discrete probability graph, includes: Construct a two-dimensional cylindrical expansion coordinate system and establish a discrete mesh matrix describing the inner wall of the rainwater pipe; Based on the coordinate mapping relationship, the spatial position coordinates of each spatial node and the defect probability value are projected onto the corresponding grid points of the discrete grid matrix; Using a spatial interpolation algorithm, smooth transition interpolation is performed on unprojected grid points based on grid points with projected defect probability values ​​to generate the discrete probability map containing the global probability distribution.

8. The intelligent identification method for rainwater pipe defects based on deep learning according to claim 1, characterized in that, Step S5, the step of processing the probability gradient features in the discrete probability graph, includes: Calculate the differential derivative of the discrete probability map and extract the extreme points that characterize the probability gradient features of the defect interface; Using the envelope of the extreme point as the initial contour, an active contour functional containing an internal tension constraint term and an external gradient attraction term is constructed as the fitting function for the closed curve. The continuous defect boundary profile is extracted by minimizing the computation to drive the active profile functional to shrink and close towards the extreme value region.

9. The intelligent identification method for rainwater pipe defects based on deep learning according to claim 1, characterized in that, Step S6, the step of calculating the enclosing area of ​​the continuous defect boundary profile, includes: Obtain the pipe geometry parameters in the three-dimensional twin model, and inversely map the continuous defect boundary contour to a three-dimensional spatial surface based on the pipe geometry parameters; The surface integral algorithm is used to perform surface integral along the mapped continuous defect boundary contour on the three-dimensional space surface to calculate the enclosed area; The estimated repair volume, including the defect morphology, is calculated by multiplying the enclosed area with a preset pipe wall depth parameter.

Images