Drainage pipe blockage disease diagnosis method coupling domain knowledge and deep learning

Abstract

The application discloses a kind of coupling field knowledge and deep learning's sewer pipe silt disease diagnosis method, the method includes: according to the momentum conservation principle of water reference unit, realize the pipe silt microscopic hydraulics model analysis based on fluid dynamics, extract pipe silt overflow characteristic field knowledge;Using discrete element analysis and computational fluid dynamics coupling numerical simulation, the parameter value of pipe silt microscopic hydraulics model is calibrated;Intelligent diagnosis model of pipe silt disease under the coupling action of field knowledge and deep small world neural network is established and the model is trained;Multi-dimensional training sample data set data acquisition mode, detect the abnormal value in sample data set, while interpolation operation is carried out to missing value;Using multi-scale attention neural network as the classifier of gradually improved generative adversarial network, improve the accuracy of pipe entrance and exit flow velocity / flow time series sample data set expansion result.

Description

Technical Field

[0001] This invention belongs to the field of integrated technology of municipal drainage pipeline disease detection and diagnosis, data processing technology and artificial intelligence technology, and relates to a method for diagnosing drainage pipeline siltation diseases by coupling domain knowledge and deep learning. Background Technology

[0002] There is a significant and complex coupling relationship between pipeline siltation and flow characteristics (inlet and outlet velocities and flow rates). To investigate the instantaneous impact of pipeline siltation on flow, existing research has proposed a force analysis method based on a stochastic initiation model of pipeline siltation. In these models, different values ​​of the initiation conditions for pipeline siltation will cause significant deviations in the calculation of flow state parameters. With the introduction of the theory of pipeline siltation initiation, existing research has established a model of the coherent interaction between pipeline siltation initiation and turbulence. However, the domain knowledge, initial conditions, and boundary conditions regarding the dynamic changes of inlet and outlet velocities / flow rates over time under the influence of pipeline siltation have not been effectively extracted. Existing research has analyzed the quantitative relationship between water hammer waves, the degree of pipeline siltation, and its length based on dynamic differential equations and experiments. However, these studies have not considered the influence of the feedback relationship between pipeline siltation and flow on pipeline flow characteristics. Therefore, how to construct a micro-hydraulic model of pipeline sedimentation and quantitatively analyze the pipeline velocity / flow control equations, boundary conditions, and initial conditions in the model is one of the key issues that urgently need to be addressed.

[0003] With the rapid development of artificial intelligence and computing technology, the degree and length of siltation can be assessed using AI-based fault detection methods. Currently, methods for fault detection and diagnosis of complex systems based on data mining technology mainly include expert systems, Bayesian networks, rough sets, Petri nets, and deep neural networks. Although scholars have expanded and improved the analysis of the flow characteristics of siltation in drainage pipes, they have not considered the multiphase flow feedback relationships between suspended mass, bedload, water flow, and upper air, as well as the influence of multi-field coupling such as gravity field, stress field, and flow field on the flow characteristics of pipes. Meanwhile, the loss function optimization process in existing intelligent diagnostic models for pipe siltation lacks control equations and boundary constraints guided by knowledge of the fluid-structure interaction domain of pipe siltation, leading to partial distortion in the output results. Therefore, how to combine knowledge of drainage pipe siltation, flow characteristic index data collected by monitoring systems, and deep learning to construct an intelligent diagnostic model for drainage pipe siltation under the coupling effect of domain knowledge and deep learning is another key problem that this research urgently needs to solve.

[0004] In summary, how to construct an intelligent diagnostic model for drainage pipe siltation from three aspects—a micro-hydraulic model of drainage pipe siltation, "outlier detection-optimal interpolation-sample expansion" of pipeline inlet / outlet velocity / flow sequence datasets, and an intelligent diagnostic algorithm based on the coupling effect of domain knowledge and deep learning—is an urgent issue that needs to be addressed. Summary of the Invention

[0005] To address the above problems, the purpose of this invention is to provide a diagnostic method for drainage pipe siltation based on three aspects: microscopic hydraulic model analysis of drainage pipe siltation, DEM-CFD fluid-structure interaction numerical simulation, error correction of the loss function of a deep small-world neural network based on knowledge of pipe siltation, and "outlier detection-missing value interpolation-expansion" of the sample dataset.

[0006] The technical solution adopted in this invention is as follows: S1: Based on the principle of momentum conservation, the principle of incompressibility of water flow and the principle of volume conservation of the water reference unit in the drainage pipe, the Reynolds time-averaged motion equation and the boundary conditions of the pipe inlet and outlet of the pipe sedimentation micro-hydraulic model are obtained.

[0007] S2: By discretizing the differential equation and applying the principle that the flow rate of the pipe remains constant under the condition of no overflow, the equation for the conservation of angular momentum of siltation in the drainage pipe is obtained.

[0008] S3: Parameters of a multi-field coupled transient hydraulic model for pipeline siltation are calibrated through numerical simulation coupled with discrete element analysis and computational fluid dynamics;

[0009] S4: The training and test sample datasets were obtained using a multi-dimensional data acquisition method that combines full-scale testing, coupled fluid dynamics numerical simulation, and regular on-site monitoring.

[0010] S5: Outlier detection and missing value interpolation for training and testing sample datasets. Outliers in the sample dataset are detected by the local outlier detection method. At the same time, the missing values ​​are interpolated using the inverse squared regression distance method and the regression co-kriging method. The accuracy of the two interpolation methods is compared and analyzed by cross-validation, and the method with the higher accuracy is selected for interpolation.

[0011] S6: Using a multi-scale attention neural network as a classifier for progressively improving generative adversarial networks, the data of the collected sample dataset is augmented to improve the accuracy of the augmented results of the pipeline inlet / outlet velocity / flow time series sample dataset.

[0012] S7: Using small-world neural networks, restricted Boltzmann machine networks, and domain knowledge embedding methods, an intelligent diagnostic model for drainage pipe siltation is established under the coupling effect of domain knowledge and deep small-world neural networks, and the intelligent diagnostic model is trained.

[0013] S8: Use the training sample dataset to repeatedly train the weight parameters in the intelligent diagnostic model. When the intelligent diagnostic model is trained to maturity, use the input test sample dataset to dynamically output the pipeline siltation diagnosis results.

[0014] Furthermore, the Reynolds time-averaged equation of motion and the boundary conditions at the pipe inlet and outlet in step S1 are given in formulas (1), (2), and (3):

[0015]

[0016] u(A,t)=v A (t) (2)

[0017] u(B,t)=v B (t) (3)

[0018] Where u represents the component of the average flow velocity of the water in the pipe along the x-axis in the three-dimensional coordinate system, t represents a specific moment in the time dimension, x represents the coordinate axis parallel to the pipe direction, g represents the gravitational acceleration of the water flow in the pipe, and α represents the angle between the pipe and the standard horizontal plane. This represents the real-time fluctuating velocity of component u along the x-axis in a three-dimensional coordinate system, and ζ is the turbulent kinematic viscosity coefficient, v t Let μ represent the turbulent kinetic energy per unit mass of water flow, and μ represent the dynamic viscosity coefficient. Let u(A,t) and u(B,t) represent the Laplace function in three-dimensional space, respectively, and let v represent the average flow velocities at the pipe inlet and outlet at time t. A (t) and v B (t) represent the flow velocities at the pipe inlet and outlet at time t, as measured by the flow meter.

[0019] Furthermore, in step S2, the equation for the conservation of angular momentum of the siltation in the drainage pipe, and the theoretical flow velocity v at the pipe outlet... BT The specific expressions for (t+t1) and the time t1 required for the water flow to move from pipe inlet A to pipe outlet B are given in formulas (4), (5) and (6):

[0020]

[0021]

[0022]

[0023] Where ρ is the sediment-laden flow density, Q(t) represents the instantaneous flow rate at time t passing the target location in the pipe, Δt represents the infinitesimal time interval, and F... fw and F fs The frictional resistances of fluid-structure interaction in non-silting and silting pipe sections are respectively represented by l. p This represents the total length of the pipeline from inlet A to outlet B. c η3 represents the length of the pipe where siltation has occurred, and η3 is the kinetic energy conversion efficiency coefficient of the non-silted pipe section. c The thickness of the silt deposit in the pipe is represented by t1, which represents the time required for water to travel from pipe inlet A to pipe outlet B. BT (t+t1) represents the theoretical flow velocity of the water at pipe outlet B, v A (t) represents the flow velocity of water at pipe inlet A at time t, and η1 and η2 represent the friction coefficients of the non-silted pipe and the silted pipe, respectively.

[0024] Furthermore, in step S3, the parameters of the multi-field coupled transient hydraulic model of pipeline siltation are calibrated through finite element analysis and full-scale tests, while verifying and eliminating outliers collected in the full-scale tests.

[0025] Furthermore, the outlier detection and missing value interpolation of the training and testing sample datasets in step S5 specifically include the following steps:

[0026] Step S51: Calculate the k-th smallest distance dk(p) of the pipeline inlet / outlet velocity / flow rate sequence to be detected, which is defined as dk(p). k (p) = d(p,o), where the distance is in the time dimension, satisfying:

[0027] (1) There are at least k points in set C that do not include p. For any point q, q∈C{q≠p}, d(p,q)≤d(p,o);

[0028] (2) There are at most k-1 points in the set C, excluding p. For any point q, if q∈C{q≠p}, then d(p,q)<d(p,o).

[0029] Step S52: Calculate the neighborhood N of the k-th smallest distance of the index data p. k (p), the number of neighborhood points that satisfy the k-th smallest distance of index data p |N k (p)|≥k;

[0030] Step S53: Calculate the k-th smallest reachable distance from indicator data o to indicator data p.

[0031] d k (p,o)=max{d k(o),d(p,o)};

[0032] Step S54: Calculate the local reachability density of the index data p

[0033] Step S55: Calculate the local outlier of the index data p

[0034] Step S56: For the sample dataset containing outliers, the Local Outlier Detection Algorithm is used to calculate an LOF factor for each sample data. Based on the calculated LOF factor and the set threshold, it is determined whether the sample is an outlier sample. Samples represented by LOF factors exceeding the set threshold are classified as outlier samples, and those below the threshold are normal samples.

[0035] Step S57: The accuracy of the two interpolation methods, Inverse Squared Regression (IDS) and Co-Kriging (CK), is compared and analyzed using cross-validation. The method with higher accuracy is selected for interpolation, as shown in the following formula:

[0036]

[0037] Among them, V opt (x i ) represents the optimal interpolation for the outlier value of the i-th collected indicator data, x i-1 x i x i+1 Let ME represent the actual values ​​of the (i-1), i, and i+1th collected indicator data, respectively. IDS and RMSE IDS These represent the root mean square error and the average error of the IDS, respectively. CK and RMSE CK V represents the root mean square error and the average error of CK, respectively. IDS (x i ) and V CK (x i ) represent the interpolations obtained using the IDS and CK methods, respectively.

[0038] Furthermore, in step S6, the collected sample dataset is expanded, specifically through the following steps:

[0039] Step S61: Use convolution instead of deconvolution to construct a generator for the generative adversarial network to gradually improve the generator structure, so as to ensure that the generator structure in the generative adversarial network has universality when processing time series data in various fields.

[0040] Step S62: Introduce gradient penalty theory. The kernel size of each convolutional layer in the generator is fixed at 3×3. In the discriminator, the kernel size of the convolutional layer in each multi-scale convolutional unit is fixed at 3×3, 6×6 and 12×12.

[0041] Step S63: The discriminator classifies true and false data by determining whether the probability that the output augmented data sample dataset matches the real data sample dataset is greater than the threshold of 0.5.

[0042] Furthermore, in step S7, an intelligent diagnostic model for drainage pipe siltation is developed, and the model is trained. Specific steps include:

[0043] Step S71: Using the sequence of flow velocity / flow rate changes over time at the pipe inlet and outlet as input, use a sliding window with a duration of 8 seconds to extract all sample data to obtain the mean subsequence under the sliding window; construct all subsequences into a training sample dataset, and add a pipe siltation status label to identify the training samples of the silted pipe section.

[0044] Step S72: Based on formulas (1)-(6), introduce a penalty term for the governing equation. and Initial condition penalty term MSE PIC and boundary condition penalty term MSE PBC The specific calculation methods for the penalty terms of the control equation, initial condition penalty terms, and boundary condition penalty terms are given in formulas (8), (9), (10), and (11):

[0045]

[0046]

[0047]

[0048]

[0049] Where, N PGE N PIC and N PBC Q represents the number of velocity / flow rate data sequences at the pipe inlet and outlet, the number of boundary conditions, and the number of initial conditions, respectively. A (t) and Q B (t) represents the instantaneous flow rate through pipe inlet A and outlet B at time t, respectively. i l represents the instantaneous flow velocity at the same location in the ith time period within the same algorithm training cycle. c,i and s c,iLet u represent the diagnostic results of the pipe siltation length and thickness at the end of the (i-1)th round of training. i (A,t) and u i (B, t+t1) represent the initial and boundary conditions derived from the diagnostic results of the current training round of the intelligent algorithm, respectively. t Let μ represent the turbulent kinetic energy per unit mass of water flow, and v represent the dynamic viscosity coefficient. A (t) represents the flow velocity of the water at pipe inlet A at time t, v BT (t+t1) represents the theoretical flow velocity of the water at pipe outlet B. This represents the penalty term in the governing equation derived from formula (1). The penalty term of the governing equation derived from formula (4), MSE PIC MSE represents the initial condition penalty term. PBC Indicates the boundary condition penalty term;

[0050] Step S73: Based on the above analysis, the penalty terms of the control equation, boundary conditions, and initial conditions are introduced into the classic loss function of deep small-world neural networks to obtain the domain knowledge-guided loss function error correction formula, as shown in formula (11):

[0051]

[0052] Where, η PGE η PIC η PBC and η DATA Let MSE represent the mean square error weights of the control equations obtained from formula (1) and formula (4) in the domain knowledge-guided loss function error correction formula, respectively, as well as the initial and boundary conditions. DATA This represents the mean squared error between the pipe congestion diagnosis results of a deep small-world neural network before the introduction of domain knowledge and the input sample labels. This represents the penalty term in the governing equation derived from formula (1). The penalty term of the governing equation derived from formula (4), MSE PIC MSE represents the initial condition penalty term. PBC L represents the boundary condition penalty term. c The error correction formula for the loss function guided by domain knowledge;

[0053] Step S74: Use an unlabeled dataset to pre-train a multilayer restricted Boltzmann machine network in a deep small-world neural network (DSWNN) that embeds domain knowledge. After obtaining sufficient feature representations, record the network parameters, including the weights and biases of neurons.

[0054] Step S75: Randomly add edges to the multi-layer RBM network according to the domain knowledge and the random edge addition probability p of the deep small-world neural network (DSWNN), and add a fully connected classification layer at the top layer to transform it into a deep small-world neural network with multi-classification function; supervised training of the network is performed again using supervised learning, and the parameters are fine-tuned using the backpropagation algorithm until the training is completed, and the trained domain knowledge and deep small-world neural network (DSWNN) classifier model is obtained;

[0055] Step S76: Input the test sample dataset of the inlet and outlet flow velocity and flow rate sequences of the pipeline into the pre-trained domain knowledge and deep small-world neural network (DSWNN) classifier model that embeds domain knowledge of pipeline siltation, and obtain the output results of pipeline siltation degree and siltation length.

[0056] The beneficial effects of this invention are: This invention provides a basis and means for the accurate and efficient diagnosis of siltation problems in drainage pipes. The method is a complex set of approaches involving fluid dynamics theory, multi-dimensional experimental methods, deep learning technology, knowledge embedding technology, and dataset expansion and optimization techniques. Attached Figure Description

[0057] Figure 1 A detailed flowchart of the pipeline siltation diagnosis method of the present invention is shown;

[0058] Figure 2 This invention illustrates the structural diagram of a microscopic hydraulic model for pipeline siltation.

[0059] Figure 3 A schematic diagram illustrating the physical analysis of the momentum conservation equation of this invention is shown.

[0060] Figure 4 This diagram illustrates the flow chart of the DEM-CFD fluid-structure interaction numerical simulation of pipe siltation in this invention.

[0061] Figure 5 A schematic diagram of the deep small-world neural network structure of the present invention is shown.

[0062] Figure 6 This diagram illustrates the error correction mechanism of the loss function in the field of pipeline siltation knowledge according to the present invention.

[0063] Figure 7 This diagram illustrates the sample dataset expansion process of the present invention. Detailed Implementation

[0064] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, and to make the above and other objects, features and advantages of the present invention more easily understood, specific embodiments of the present invention are listed below.

[0065] like Figure 1 As shown, the present invention discloses an intelligent diagnostic method for drainage pipe siltation defects based on the coupling effect of domain knowledge and deep learning. The steps are as follows:

[0066] Step 1: Establish a micro-hydraulic model of siltation in drainage pipes based on fluid dynamics, and construct the Reynolds time-averaged equation of motion;

[0067] Step 2: Establish the conservation equation of angular momentum for siltation in drainage pipes, and extract the governing equations and boundary conditions from the domain knowledge of pipe siltation flow characteristics;

[0068] Step 3: Use DEM-CFD fluid-structure interaction numerical simulation to calibrate the parameter values ​​of the microscopic hydraulic model for pipeline sedimentation.

[0069] Step 4: Establish an intelligent diagnostic model for drainage pipe siltation under the coupling effect of domain knowledge and DSWNN, and introduce a loss function error correction mechanism based on domain knowledge of pipe siltation.

[0070] Step 5: The training sample dataset is collected using a combination of full-scale testing, discrete element analysis and computational fluid dynamics (DEM-CFD) numerical simulation, and regular on-site monitoring.

[0071] Step 6: Perform outlier detection and missing value interpolation on the training and testing sample datasets;

[0072] Step 7: Use a multi-scale attention neural network as a classifier for progressively improving generative adversarial networks (SIGAN) to expand the training and test sample datasets.

[0073] Step 8: Repeatedly train the intelligent diagnostic model for pipeline siltation and output diagnostic results using a test sample dataset.

[0074] like Figure 2 As shown, based on the principles of momentum conservation, incompressibility, and volume conservation of water reference elements in drainage pipes, a microscopic hydraulic model of pipe sedimentation in this invention is constructed. The Reynolds time-averaged equation of motion and the pipe inlet / outlet boundary conditions for this microscopic hydraulic model of pipe sedimentation are obtained as follows:

[0075]

[0076] u(A,t)=v A (t) (2)

[0077] u(B,t)=v B (t) (3)

[0078] Where u represents the component of the average flow velocity of the water in the pipe along the x-axis (parallel to the pipe) in the three-dimensional coordinate system, and t represents a specific moment in the time dimension. Let represent the Laplace function in three-dimensional space, g be the gravitational acceleration of the water flow, x be the coordinate axis parallel to the pipe direction, α be the angle between the pipe and the standard horizontal plane, u(A,t) and u(B,t) be the average flow velocities at the pipe inlet and outlet at time t, respectively, and v A (t) and v B (t) represent the actual flow rates at the pipe inlet and outlet, respectively, as measured by the flow meter. This represents the real-time fluctuating velocity of component u along the x-axis in a three-dimensional coordinate system, and ζ is the turbulent kinematic viscosity coefficient, v t Let μ represent the turbulent kinetic energy per unit mass of water flow, and μ represent the dynamic viscosity coefficient.

[0079] The physical analysis process of the momentum conservation equation of this invention is as follows: Figure 3 As shown. The equation for the conservation of angular momentum in a drainage pipe can be obtained by discretizing the differential equation and solving it using the principle of constant flow rate in the absence of overflow. The equation for the conservation of angular momentum and the theoretical flow velocity v at the pipe outlet are also relevant. BT The specific expressions for (t+t1) and the time t1 required for the water to move from pipe inlet A to pipe outlet B are as follows:

[0080]

[0081]

[0082]

[0083] Where ρ is the sediment-laden flow density, Q(t) represents the instantaneous flow rate at time t passing the target position in the pipe, Δt represents the infinitesimal time interval, g is the gravitational acceleration of the water flow, α represents the angle between the pipe and the standard horizontal plane, and F... fw and F fs Let l represent the frictional resistance of fluid-structure interaction in non-silting and silting pipe sections, respectively. p This represents the total length of the pipeline from inlet A to outlet B. c η3 represents the length of the pipe where siltation has occurred, and η3 is the kinetic energy conversion efficiency coefficient of the non-silted pipe section. c The thickness of the silt deposit in the pipe is represented by t1, which represents the time required for water to travel from pipe inlet A to pipe outlet B. BT (t+t1) represents the theoretical flow velocity of the water at pipe outlet B, v A (t) represents the flow velocity of water at pipe inlet A at time t, and η1 and η2 represent the friction coefficients of the non-silted pipe and the silted pipe, respectively.

[0084] The process of DEM-CFD fluid-structure interaction numerical simulation of pipeline siltation in this invention is as follows: Figure 4 As shown, a combination of Fluid Numerical Simulation Software (FLUENT) and Particulate Numerical Simulation Software (EDEM) is used to achieve DEM-CFD coupling. EDEM is responsible for generating discrete particles and setting boundary conditions, and is linked to FLUENT software via a multiphase flow API secondary development programming interface. Furthermore, a control unit submodule for the physical properties of the pipeline fluid and sediment particles is nested within FLUENT. After each time step of fluid-structure interaction calculation in FLUENT, the flow field data is imported into EDEM to calculate the state equation of particle phase motion and displacement within the corresponding time step. At the end of the calculation round, FLUENT determines the changes in the velocity / flow rate sequence at the pipeline inlet and outlet based on the volume fraction and the coupling calculation results between the two phases.

[0085] This invention utilizes a deep small-world neural network ( Figure 5 ), Error correction mechanism for loss function in the field of pipeline siltation knowledge ( Figure 6 ), outlier detection and interpolation of sample datasets, and sample dataset augmentation ( Figure 7 Methods such as [list of methods] are used to improve the accuracy of intelligent diagnosis of drainage pipe siltation. Each module has been improved by modifying the source code. The specific steps are as follows:

[0086] S1: Build a deep small-world neural network model. The specific steps are as follows:

[0087] Step S11: Use an unlabeled dataset to pre-train a multilayer restricted Boltzmann machine (RBM) network in a deep small-world neural network. After obtaining sufficient feature representations, record the network parameters, including the weights and biases of the neurons.

[0088] Step S12: Randomly add edges to the multi-layer RBM network according to the random edge addition probability p of the deep small-world neural network, and add a fully connected classification layer at the top layer to transform it into a deep small-world neural network with multi-classification function; supervised training of the network is performed again using supervised learning, and the parameters are fine-tuned using the backpropagation (BP) algorithm until training is completed, and a trained deep small-world neural network classifier model is obtained.

[0089] S2: Introduce knowledge from the field of pipeline siltation (governing equations, boundary conditions, initial conditions) into the error correction of the loss function of the deep small-world neural network to improve the accuracy and robustness of the algorithm. The specific steps are as follows:

[0090] Step S21: Using the sequence of flow velocity / flow rate changes over time at the pipe inlet and outlet as input, use a sliding window with a duration of 8 seconds to extract all sample data to obtain the mean subsequence under the sliding window; construct all subsequences into a training sample dataset, and add a pipe siltation status label to identify the training samples of the silted pipe section.

[0091] Step S22: As Figure 6 As shown, based on formulas (1)-(6), a penalty term is introduced into the governing equation ( and Initial condition penalty term (MSE) PIC ) and boundary condition penalty term (MSE) PBC These penalty terms are used to correct errors in the loss function of deep small-world neural networks. The specific calculation method for these penalty terms is as follows:

[0092]

[0093]

[0094]

[0095]

[0096] Where, N PGE N PIC and N PBC Q represents the number of velocity / flow rate data sequences at the pipe inlet and outlet, the number of boundary conditions, and the number of initial conditions, respectively. A (t) and Q B (t) represents the instantaneous flow rate through pipe inlet A and outlet B at time t, respectively. i l represents the instantaneous flow velocity at the same location in the ith time period within the same algorithm training cycle. c,i and s c,i Let u represent the diagnostic results of the pipe siltation length and thickness at the end of the (i-1)th round of training. i (A,t) and u i (B, t+t1) represent the initial and boundary conditions derived from the diagnostic results of the current training round of the intelligent algorithm, respectively. t Let μ represent the turbulent kinetic energy per unit mass of water flow, and v represent the dynamic viscosity coefficient. A (t) represents the flow velocity of the water at pipe inlet A at time t, v BT (t+t1) represents the theoretical flow velocity of the water at pipe outlet B. This represents the penalty term in the governing equation derived from formula (1). The penalty term of the governing equation derived from formula (4), MSE PICMSE represents the initial condition penalty term. PBC This indicates the boundary condition penalty term.

[0097] Step S23: Based on the above analysis, penalty terms for the control equation, boundary conditions, and initial conditions are introduced into the classic deep small-world neural network loss function to obtain the domain knowledge-guided loss function error correction formula, as shown below:

[0098]

[0099] Where, η PGE η PIC η PBC and η DATA Let MSE represent the mean square error weights of the control equations obtained from formula (1) and formula (4) in the domain knowledge-guided loss function error correction formula, respectively, as well as the initial and boundary conditions. DATA This represents the mean squared error between the pipe congestion diagnosis results of a deep small-world neural network before the introduction of domain knowledge and the input sample labels. This represents the penalty term in the governing equation derived from formula (1). The penalty term of the governing equation derived from formula (4), MSE PIC MSE represents the initial condition penalty term. PBC L represents the boundary condition penalty term. c This represents the error correction formula for the loss function guided by domain knowledge.

[0100] S3: Outlier detection and interpolation of the sample dataset, the specific steps are as follows:

[0101] Step S31: Calculate the k-th smallest distance d of the index to be detected (pipe inlet / outlet velocity / flow rate sequence). k (p), which is defined as d k (p) = d(p,o), where the distance is in the time dimension, satisfying:

[0102] (1) In set C, there are at least k points excluding p. For any point q, satisfying q∈C{q≠p}, we have d(p,q)≤d(p,o).

[0103] (2) There are at most k-1 points in the set C, excluding p. For any point q, if q∈C{q≠p}, then d(p,q)<d(p,o).

[0104] Step S32: Calculate the neighborhood N of the k-th smallest distance of the index data p. k (p), the number of neighborhood points that satisfy the k-th smallest distance of index data p |N k (p)|≥k;

[0105] Step S33: Calculate the k-th smallest reachable distance from index data o to data p.

[0106] d k (p,o)=max{d k (o),d(p,o)};

[0107] Step S34: Calculate the local reachability density of the index data p

[0108] Step S35: Calculate the local outlier of the index data p

[0109] Step S36: For the sample dataset containing outliers, the Local Outlier Detection Algorithm is used to calculate an LOF factor for each sample data. Based on the calculated LOF factor and the set threshold, it is determined whether the sample is an outlier sample. Samples represented by LOF factors exceeding the set threshold are classified as outlier samples, and those below the threshold are normal samples.

[0110] Step S37: The accuracy of the two interpolation methods, Inverse Squared Regression (IDS) and Co-Kriging (CK), is compared and analyzed using cross-validation. The method with higher accuracy is selected for interpolation, as shown in the following formula:

[0111]

[0112] Among them, V opt (x i ) represents the optimal interpolation for the outlier value of the i-th collected indicator data, x i-1 x i x i+1 Let ME represent the actual values ​​of the (i-1), i, and i+1th collected indicator data, respectively. IDS and RMSE IDS These represent the root mean square error and the average error of the IDS, respectively. CK and RMSE CK V represents the root mean square error and the average error of CK, respectively. IDS (x i ) and V CK (x i ) represent the interpolations obtained using the IDS and CK methods, respectively.

[0113] S4: Expanding the training and testing sample datasets, the specific steps are as follows:

[0114] Step S41: Use convolution instead of deconvolution to construct a generator for progressively improving the generative adversarial network (SIGAN) to ensure that the generator structure in the generative adversarial network has universality when processing time series data in various fields;

[0115] Step S42: Introduce gradient penalty theory. The kernel size of each convolutional layer in the generator is fixed at 3×3. In the discriminator, the kernel size of the convolutional layer in each multi-scale convolutional unit is fixed at 3×3, 6×6 and 12×12.

[0116] Step S43: The discriminator classifies true and false data by determining whether the probability that the output augmented data sample dataset matches the real data sample dataset is greater than the threshold of 0.5.

[0117] S5: Input the test sample dataset of the inlet and outlet flow velocity and flow rate sequences of the pipeline into a pre-trained deep small-world neural network classifier model that incorporates knowledge of pipeline siltation, and obtain the output results of pipeline siltation degree and siltation length.

Claims

1. A diagnostic method for siltation defects in drainage pipes that couples domain knowledge and deep learning, characterized in that... The specific steps are as follows: S1: Based on the principle of momentum conservation, the principle of incompressibility of water flow and the principle of volume conservation of the water reference unit in the drainage pipe, the Reynolds time-averaged motion equation and the boundary conditions of the pipe inlet and outlet are obtained for the microscopic hydraulic model of pipe sedimentation. S2: By discretizing the differential equation and applying the principle that the flow rate of the pipe remains constant under the condition of no overflow, the equation for the conservation of angular momentum of siltation in the drainage pipe is obtained. S3: Parameters of a multi-field coupled transient hydraulic model for pipe siltation are calibrated through numerical simulation coupled with discrete element analysis and computational fluid dynamics; S4: The training and test sample datasets were obtained using a multi-dimensional data acquisition method that combines full-scale testing, coupled fluid dynamics numerical simulation, and regular on-site monitoring. S5: Outlier detection and missing value interpolation for training and testing sample datasets. Outliers in the sample dataset are detected by the local outlier detection method. At the same time, the missing values ​​are interpolated using the inverse squared regression distance method and the regression co-kriging method. The accuracy of the two interpolation methods is compared and analyzed by cross-validation, and the method with the higher accuracy is selected for interpolation. S6: Using a multi-scale attention neural network as a classifier for progressively improving generative adversarial networks, the data of the collected sample dataset is augmented to improve the accuracy of the augmented results of the pipeline inlet / outlet velocity / flow time series sample dataset. S7: Using small-world neural networks, restricted Boltzmann machine networks, and domain knowledge embedding methods, an intelligent diagnostic model for drainage pipe siltation is established under the coupling effect of domain knowledge and deep small-world neural networks, and the intelligent diagnostic model is trained. S8: Use the training sample dataset to repeatedly train the weight parameters in the intelligent diagnostic model. When the intelligent diagnostic model is trained to maturity, use the input test sample dataset to dynamically output the pipeline siltation diagnosis results. The Reynolds time-averaged equation of motion and the boundary conditions at the pipe inlet and outlet in step S1 are given in formulas (1), (2), and (3): （1）； （2）； （3）； Where u represents the component of the average flow velocity of the water in the pipe on the x-axis of the three-dimensional coordinate system, t represents a specific moment in the time dimension, and x represents the coordinate axis parallel to the direction of the pipe. This represents the acceleration due to gravity of the water flow in the pipe. This indicates the angle between the pipe and the standard horizontal plane. This represents the real-time fluctuating velocity of component u along the x-axis in a three-dimensional coordinate system, and , The viscosity coefficient is the coefficient of turbulent motion. The turbulent kinetic energy per unit mass of water flow. Indicates the dynamic viscosity coefficient. Represents the Laplace function in three-dimensional space. and Let represent the average flow velocities at the pipe inlet and outlet at time t, respectively. and These represent the flow velocities of the water at pipe inlet A and outlet B at time t, as measured by the flow meter. Step S6 involves expanding the collected sample dataset, specifically including the following steps: Step S61: Use convolution instead of deconvolution to construct a generator for the generative adversarial network to gradually improve the generator structure, so as to ensure that the generator structure in the generative adversarial network has universality when processing time series data in various fields. Step S62: Introduce gradient penalty theory. The kernel size of each convolutional layer in the generator is fixed at 3×3. In the discriminator, the kernel size of the convolutional layer in each multi-scale convolutional unit is fixed at 3×3, 6×6 and 12×12. Step S63: The discriminator classifies data into true and false by determining whether the probability that the output augmented data sample dataset matches the real data sample dataset is greater than the threshold of 0.

5. The equation for the conservation of angular momentum of the siltation in the drainage pipe in step S2, and the specific expressions for the theoretical flow velocity at the pipe outlet and the time required for the water to move from the pipe inlet A to the pipe outlet B are given in formulas (4), (5), and (6): （4）； （5）； （6）； in, For the density of sediment-laden water flow, This represents the instantaneous flow rate at time t as it passes through the target location in the pipe. Represents an infinitesimally small time interval. and These represent the frictional resistance of fluid-structure interaction in non-silting and silting pipe sections, respectively. This represents the total length of the pipeline from inlet A to outlet B. Represents the length of the pipe siltation. The coefficient of water kinetic energy conversion efficiency in non-siltation pipe sections. This represents the thickness of the silt buildup in the pipe. This indicates the time required for water to travel from pipe inlet A to pipe outlet B. This represents the theoretical flow velocity of the water at pipe outlet B. This represents the flow velocity of water at pipe inlet A at time t. and These represent the friction coefficients of the non-silted and silted pipes, respectively; In step S7, an intelligent diagnostic model for drainage pipe siltation is established, and the intelligent diagnostic model is trained. Specific steps include: Step S71: Using the sequence of flow velocity / flow rate changes over time at the pipe inlet and outlet as input, use a sliding window with a duration of 8 seconds to extract all sample data to obtain the mean subsequence under the sliding window; construct all subsequences into a training sample dataset, and add a pipe siltation status label to identify the training samples of the silted pipe section. Step S72: Based on formulas (1)-(6), introduce a penalty term for the governing equation. and Initial condition penalty item and boundary condition penalty terms The specific calculation methods for the penalty terms of the control equation, initial condition penalty terms, and boundary condition penalty terms are given in formulas (8), (9), (10), and (11): （8）； （9）； （10）； （11）； in, , and These represent the number of velocity / flow data sequences at the pipe inlet and outlet, the number of boundary conditions, and the number of initial conditions, respectively. This represents the instantaneous flow velocity at the same location in the silted pipe segment within the i-th time period of the same algorithm training cycle. Let t represent the instantaneous flow rate through pipe inlet A at time t. and These represent the diagnostic results for the length and thickness of pipe siltation at the end of the (i-1)th round of training. and These represent the initial and boundary conditions derived from the diagnostic results of the current training round of the intelligent algorithm, respectively. This represents the flow velocity of water at pipe inlet A at time t. This represents the theoretical flow velocity of the water at pipe outlet B. This represents the penalty term in the governing equation derived from formula (1). This represents the penalty term in the governing equation derived from formula (4). This represents the initial condition penalty term. Indicates the boundary condition penalty term; Step S73: Based on the above analysis, the penalty term of the control equation, the boundary condition penalty term, and the initial condition penalty term are introduced into the classic deep small-world neural network loss function to obtain the domain knowledge-guided loss function error correction formula, as shown in formula (11): （11）； in, , , and Let represent the mean square error weights of the control equations obtained from formula (1) and formula (4) in the domain knowledge-guided loss function error correction formula, respectively, as well as the initial and boundary conditions. This represents the mean squared error between the pipe congestion diagnosis results of a deep small-world neural network before the introduction of domain knowledge and the input sample labels. This represents the penalty term in the governing equation derived from formula (1). This represents the penalty term in the governing equation derived from formula (4). This represents the initial condition penalty term. This represents the boundary condition penalty term. The error correction formula for the loss function guided by domain knowledge; Step S74: Use an unlabeled dataset to pre-train a multilayer restricted Boltzmann machine network in a deep small-world neural network that embeds domain knowledge. After obtaining sufficient feature representations, record the network parameters, including the weights and biases of neurons. Step S75: Randomly add edges to the multilayer restricted Boltzmann machine network based on the domain knowledge and the random edge addition probability of the deep small-world neural network, and add a fully connected classification layer at the top layer to transform it into a deep small-world neural network with multi-classification function; supervised training of the network is performed again using supervised learning, and the parameters are fine-tuned using the backpropagation algorithm until training is completed, and a well-trained domain knowledge and deep small-world neural network classifier model is obtained. Step S76: Input the test sample dataset of the inlet and outlet flow velocity and flow rate sequences of the pipeline into the pre-trained domain knowledge and deep small-world neural network classifier model that embeds domain knowledge of pipeline siltation, and obtain the output results of pipeline siltation degree and siltation length.

2. The method for diagnosing drainage pipe siltation defects by coupling domain knowledge and deep learning according to claim 1, characterized in that: In step S3, the parameters of the multi-field coupled transient hydraulic model of pipeline siltation are calibrated through finite element analysis and full-scale tests, while verifying and eliminating outliers collected in the full-scale tests.

3. The method for diagnosing drainage pipe siltation defects by coupling domain knowledge and deep learning according to claim 2, characterized in that: Step S5 involves outlier detection and missing value interpolation on the training and testing sample datasets. The specific steps include: Step S51: Calculate the k-th smallest distance of the target index, i.e., the velocity / flow rate sequence at the pipe inlet and outlet. Its definition is ,distance For the time dimension, it satisfies: (1) There are at least k points in set C, excluding p. For any point ,satisfy ,have , (2) There are at most k-1 points in set C, excluding p. For any point ,satisfy ,have ; Step S52: Calculate the neighborhood of the k-th smallest distance of the index data p. The number of neighborhood points that satisfy the k-th smallest distance of the index data p ; Step S53: Calculate the k-th smallest reachable distance from indicator data o to indicator data p. ； Step S54: Calculate the local reachability density of the index data p ; Step S55: Calculate the local outlier of the index data p ; Step S56: For the sample dataset containing outliers, the local anomaly detection algorithm is used to calculate a local anomaly factor for each sample data. Based on the calculated local anomaly factor and the set threshold, it is determined whether the sample is an abnormal sample. Samples represented by local anomaly factors that exceed the set threshold are classified as abnormal samples, and those that are less than the threshold are normal samples. Step S57: The accuracy of the two interpolation methods, the inverse squared regression distance method and the co-kriging regression method, is compared and analyzed using cross-validation. The method with the higher accuracy is selected for interpolation, as shown in the following formula: （7）； in, This represents the optimal interpolation value for the j-th collected indicator data outlier. , , These represent the actual values ​​of the (j-1), j, and j+1th collected indicator data, respectively. and Let represent the root mean square error and the average error of the inverse square method of regression distance, respectively. and Let represent the root mean square error and the average error of the regressive co-kriging method, respectively. and These represent the interpolations obtained using the inverse squared regression distance method and the regression co-kriging method, respectively.

Images