An intelligent leak detection method and device for underground pipelines

Abstract

The application discloses an underground pipeline intelligent leakage detection method and device, relates to the field of intelligent leakage detection, collects a high-frequency pressure sequence, establishes a time-temperature equivalent equation to calculate a moving factor, maps the high-frequency pressure sequence to a reduced time axis by using the moving factor, obtains a temperature equivalent pressure sequence, and is converted into a full history stress-strain memory field tensor, constructs a neural lag operator network, takes the full history stress-strain memory field tensor as input, carries out non-local memory characteristic and hysteresis loop analysis, carries out point-by-point difference processing on the high-frequency pressure sequence and the pipe material intrinsic nonlinear hysteresis response signal, extracts a fidelity fluid dynamics abnormal residual error, constructs a spatiotemporal evolution Poincare section, inputs into a convolutional neural network, and determines real leakage negative pressure waves and sensor random drift through attractor shape difference on the spatiotemporal evolution Poincare section, so that periodic false alarms caused by material rheological characteristics are eliminated.

Description

Technical Field

[0001] This invention relates to the field of intelligent leak detection, specifically to an intelligent leak detection method and device for underground pipelines. Background Technology

[0002] For leak detection in underground pipe networks of non-metallic materials such as high-density polyethylene, highly sensitive dynamic pressure sensors are widely used to capture minute transient negative pressure wave signals. In practical engineering applications, especially during the steady-state window of low flow and low noise at night, monitoring systems often capture a weak pressure transient signal with quasi-periodic characteristics. Existing signal processing techniques, usually based on linear elasticity theory or simple time-frequency threshold discrimination, often fail to define the physical properties of such signals, frequently misattributing them to random zero-point drift of the sensor, voltage fluctuations in the electronic system, or unknown background noise. This technical misjudgment leads to frequent, unreproducible phantom false alarms from the system.

[0003] The aforementioned false alarms do not stem from hardware failures in the sensors themselves, but rather from the fact that current technology overlooks the nonlinear modulation effect of the inherent rheological properties of polymer pipes on fluid dynamics signals. Unlike the instantaneous elastic response of metal pipes, polymer materials exhibit viscoelasticity and stress relaxation effects. Their deformation response depends not only on the current pressure load but also heavily on the stress loading path throughout the entire history. Under frequent pressure fluctuations during the day, the pipe wall accumulates complex stress-strain memories. When the pressure inside the pipe stabilizes at night, the pipe undergoes hysteretic creep recovery or stress relaxation. This microscopic physical retraction of the pipe wall compresses the internal fluid, generating induced pressure waves that are easily confused with leakage signals. Current technology lacks a mathematical model capable of decoupling this intrinsic nonlinear hysteretic response of the material from actual fluid leakage anomalies, making it difficult to isolate the interference components generated by the release of material memory in the fluid-structure interaction field.

[0004] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0005] To address the technical problems mentioned in the background section, this invention is proposed. This invention provides an intelligent leak detection method and apparatus for underground pipelines.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A smart leak detection method for underground pipelines, comprising the following steps:

[0008] The high-frequency pressure sequence of the target pipe section was collected. Based on the Boltzmann superposition principle, the high-frequency pressure sequence was transformed into a full-history stress-strain memory field tensor by using ambient temperature as a thermorheological adjustment factor.

[0009] A neural hysteresis operator network is constructed, and the full-history stress-strain memory field tensor is used as input to perform nonlocal memory characteristics and hysteresis loop analysis to obtain the intrinsic nonlinear hysteresis response signal of the pipe.

[0010] The high-frequency pressure sequence and the intrinsic nonlinear hysteresis response signal of the pipe are subjected to point-by-point differential processing to remove non-stationary background fluctuations, avoid the modulation interference of material nonlinearity on fluid signals, and extract the fidelity fluid dynamics anomaly residual.

[0011] The fidelity fluid dynamics anomaly residual is reconstructed into a high-dimensional phase space trajectory, and a spatiotemporal evolution Poincaré cross section is constructed. This cross section is then input into a convolutional neural network. By analyzing the attractor morphology differences on the spatiotemporal evolution Poincaré cross section, the real leakage negative pressure wave and the sensor random drift are determined.

[0012] Furthermore, the analysis steps of the full-history stress-strain memory field tensor include:

[0013] High-frequency pressure sequences of the target pipe section are collected. Based on the viscoelastic rheological response of the pipe, stress components are processed by cone screening and time weighting to construct the historical vector of pipe deformation trajectory.

[0014] Establish a time-temperature equivalent equation to calculate the shift factor;

[0015] By using a shift factor, the high-frequency pressure sequence on the physical time axis is mapped to the reduced time axis to obtain the de-temperature equivalent pressure sequence;

[0016] Based on the inversion of the material relaxation time spectrum using the de-temperature equivalent pressure sequence, the modulus intensity of each material relaxation time spectrum is arranged according to the relaxation time distribution to construct the material relaxation spectral density matrix.

[0017] The deformation trajectory history vector of the pipe is subjected to time-series difference processing to obtain the deformation increment sequence. The deformation increment sequence is then subjected to tensor product operation with the relaxation spectral density matrix to generate the full history stress-strain memory field tensor.

[0018] Furthermore, the analysis steps for constructing the pipe deformation trajectory history vector include:

[0019] Establish a pressure-time two-dimensional projection plane, take the hydrostatic pressure balance baseline as the physical reference axis, and map discrete pressure sampling points as a set of fluctuation amplitude points relative to the hydrostatic pressure balance baseline;

[0020] The dynamic geometric chord truncation method is executed to connect adjacent fluctuation amplitude point sets to form a pressure evolution chord vector. Based on the slope of the pressure evolution chord vector and the inherent elastic-viscosity critical threshold of the pipe, the evolution chord vector is marked in geometric space as a transient elastic impact vector cluster and a steady-state viscosity creep vector cluster.

[0021] Furthermore, the analysis step of constructing the pipe deformation trajectory history vector also includes:

[0022] On the pressure-time two-dimensional projection plane, with the positive time axis as the central axis and the inherent elastic-viscosity critical threshold of the pipe as the semi-vertical angle, an open triangular region in the reverse time axis direction is constructed to obtain the viscosity-dominated memory decay cone.

[0023] Using the viscous dominant memory decay cone as a geometric constraint, orthogonal elimination or high-damping attenuation is performed on the transient elastic impact vector cluster, while preserving and enhancing the projection components of the steady-state viscous creep vector cluster on the cone axis. Based on the time span, weighted aggregation is performed to construct the pressure evolution string feature synthesis vector.

[0024] The initial elastic modulus of the pipe material is obtained and its reciprocal is calculated to obtain the instantaneous compliance coefficient. The instantaneous compliance coefficient is used as a mapping factor to perform a scalar product transformation on the pressure evolution chord feature synthesis vector to obtain the pipe deformation trajectory history vector.

[0025] Furthermore, the steps of the nonlocal memory characteristics and hysteresis loop analysis include:

[0026] The network input layer receives the full-history stress-strain memory field tensor, and uses a multi-head attention mechanism to weight the influence of different historical moments on the current state to obtain a weighted fusion feature vector;

[0027] The weighted fused eigenvectors are multiplied by the full-history stress-strain memory field tensor to obtain the equivalent relaxed state vectors.

[0028] Map the equivalent relaxed state vector space to a scalar equivalent driving force;

[0029] Based on the scalar equivalent driving force, parallel PI neurons are used, with each neuron acting as a stopping operator. The threshold distribution function of the operators is optimized through a deep learning backpropagation algorithm to fit the nonlinear hysteresis of the pipe and obtain the intrinsic nonlinear hysteresis response signal of the pipe.

[0030] Furthermore, the analytical steps for stripping away non-stationary background fluctuations include:

[0031] The high-frequency pressure sequence is mapped point by point to the intrinsic nonlinear hysteresis response signal of the pipe on the time axis;

[0032] The preliminary residual is obtained by performing a subtraction operation on the subsequent intrinsic nonlinear hysteresis response signal of the pipe, and the residual is decomposed into intrinsic mode functions by performing Hilbert-Huang transform on the residual.

[0033] Based on the Shannon entropy criterion, the slow-varying characteristic IMF component of the pipe material is removed from the intrinsic mode function, and the pure fluid dynamics anomaly residual is reconstructed.

[0034] Furthermore, the analytical steps for avoiding the modulation interference of material nonlinearity on fluid signals include:

[0035] The operator density distribution features of PI neurons and the multi-scale entropy features of each IMF component are extracted to obtain the first feature vector and the second feature vector.

[0036] The first and second feature vectors are concatenated to obtain a high-dimensional joint feature vector, and a hysteresis-entropy conjugate feature mapping space is constructed.

[0037] Furthermore, the analytical steps for avoiding the modulation interference of material nonlinearity on fluid signals also include:

[0038] In the hysteresis-entropy conjugate feature mapping space, the high-dimensional joint feature vector is decomposed into an orthogonal subspace to construct the material memory principal plane and the fluid disturbance orthogonal complementary plane, and the signal component vector is decoupled to obtain the material memory component vector and the fluid disturbance component vector.

[0039] The fluid dynamics confidence coefficient is obtained by calculating the ratio of the L2 norm of the fluid disturbance component vector to the sum of the L2 norms of the material memory component vector and the fluid disturbance component vector.

[0040] The pure fluid dynamics anomaly residuals are reconstructed using soft threshold weighting using the fluid dynamics confidence coefficients to synthesize authentic fluid dynamics anomaly residuals.

[0041] Furthermore, the analysis steps for determining the true leakage negative pressure wave and the random drift of the sensor also include:

[0042] The Takens embedding theorem is used to reconstruct the high-dimensional phase space trajectory from the fidelity fluid dynamics anomaly residuals.

[0043] By extracting a Poincaré section in phase space, the continuous dynamic trajectory is transformed into a discrete two-dimensional point cloud distribution map;

[0044] The discrete two-dimensional point cloud distribution map is input into a convolutional neural network to perform steady-state attractor pattern recognition and chaotic divergence trajectory analysis to determine the real leakage negative pressure wave and sensor random drift. Among them, steady-state attractor pattern recognition includes convergent point or closed loop structure analysis, and chaotic divergence trajectory analysis includes divergent singular attractor or non-closed fracture structure analysis.

[0045] An intelligent leak detection device for underground pipelines includes:

[0046] The memory conversion module is used to acquire the high-frequency pressure sequence of the target pipe section. Based on the Boltzmann superposition principle, the high-frequency pressure sequence is converted into the full-history stress-strain memory field tensor by using the ambient temperature as a thermorheological adjustment factor.

[0047] The hysteresis response module is used to construct a neural hysteresis operator network, taking the full-history stress-strain memory field tensor as input, performing nonlocal memory characteristics and hysteresis loop analysis to obtain the intrinsic nonlinear hysteresis response signal of the pipe.

[0048] The abnormal residual module is used to perform point-by-point differential processing on the high-frequency pressure sequence and the intrinsic nonlinear hysteresis response signal of the pipe, remove non-stationary background fluctuations, avoid the modulation interference of material nonlinearity on the fluid signal, and extract the fidelity fluid dynamic abnormal residual.

[0049] The leakage determination module is used to reconstruct the fidelity fluid dynamics anomaly residual into a high-dimensional phase space trajectory, construct a spatiotemporal evolution Poincaré cross section, input it into a convolutional neural network, and determine the real leakage negative pressure wave and sensor random drift through the attractor morphology differences on the spatiotemporal evolution Poincaré cross section.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] This invention acquires high-frequency pressure sequences of target pipe sections and, based on the Boltzmann superposition principle, transforms these sequences into a full-history stress-strain memory field tensor by using ambient temperature as a thermorheological adjustment factor. A neural hysteresis operator network is then constructed, and this full-history stress-strain memory field tensor is used as input for nonlocal memory characteristic and hysteresis loop analysis to obtain the intrinsic nonlinear hysteresis response signal of the pipe material. Point-by-point differential processing is then performed on the high-frequency pressure sequence and the intrinsic nonlinear hysteresis response signal of the pipe material to remove nonstationary background fluctuations, avoid the modulation interference of material nonlinearity on the fluid signal, and extract faithful fluid dynamics anomaly residuals. By constructing a full-history stress-strain memory field tensor based on the Boltzmann superposition principle and combining it with a neural hysteresis operator network for nonlocal memory characteristic analysis, this invention effectively solves the problem of ghost-like false alarms caused by viscoelastic rheology in polymer pipes. Unlike traditional detection methods that ignore the physical properties of pipes, this invention can accurately fit the intrinsic nonlinear hysteresis response signal of pipes under complex pressure histories, quantifying and isolating non-stationary background fluctuations induced by pipe wall creep recovery or stress relaxation at a physical level. This invention avoids the modulation interference of material nonlinearity on fluid signals, ensuring that abnormal signals captured during the nighttime steady-state window originate from genuine abrupt changes in fluid dynamics, rather than stress memory released by the pipe itself, thus fundamentally eliminating periodic false alarms caused by material rheological properties.

[0052] This invention reconstructs the high-dimensional phase space trajectory of the fidelity fluid dynamics anomaly residuals, constructs a spatiotemporal evolution Poincaré cross section, and inputs it into a convolutional neural network. By analyzing the attractor morphology differences on the spatiotemporal evolution Poincaré cross section, it determines the difference between the actual leakage negative pressure wave and the sensor's random drift. Through hysteresis-entropy conjugate feature mapping space and spatiotemporal evolution Poincaré cross section analysis techniques, it achieves deep vector decoupling and topological morphology recognition of weak leakage signals. By performing orthogonal subspace decomposition in the feature space, this invention can accurately separate the mixed signal into material memory components and fluid disturbance components, and uses a convolutional neural network to identify attractor morphology differences on the Poincaré cross section, such as steady-state attractors and chaotic divergent singular attractors. This judgment logic based on nonlinear dynamic topology breaks through the limitations of traditional time-domain amplitude or frequency-domain energy thresholds, and can sensitively distinguish between the negative pressure wave caused by actual leakage and the noise generated by sensor random drift. It achieves reliable extraction and qualitative analysis of fidelity fluid dynamics anomaly residuals even in extremely low signal-to-noise ratio environments. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to show the main idea of ​​the present invention.

[0054] Figure 1 A flowchart of a method for intelligent leak detection of underground pipelines;

[0055] Figure 2 This is a system block diagram of an intelligent leak detection device for underground pipelines;

[0056] Figure 3 A schematic diagram of the vector physical attribute labeling process based on the geometric secant method;

[0057] Figure 4 This is a schematic diagram of frustoselective projection based on physical partition marking. Detailed Implementation

[0058] 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 also within the scope of protection of the present invention.

[0059] Example 1

[0060] like Figure 1 As shown, an intelligent leak detection method for underground pipelines is proposed.

[0061] Step S100: Collect the high-frequency pressure sequence of the target pipe section. Based on the Boltzmann superposition principle, the high-frequency pressure sequence is converted into the full-history stress-strain memory field tensor by using the ambient temperature as a thermorheological adjustment factor.

[0062] The method for constructing the full-history stress-strain memory tensor specifically includes:

[0063] Step S101: Collect the high-frequency pressure sequence of the target pipe section, and construct the historical vector of pipe deformation trajectory based on the viscoelastic rheological response law of the pipe material, using cone screening and time-weighted processing of stress components.

[0064] The target pipe section refers to a specific pipeline area to be assessed for fatigue life or leak detection. This area is typically selected as a critical node section in a water supply network that experiences severe pressure fluctuations, has a high risk of geological subsidence, or has a long service life. Examples include pump station outlet sections, trunk pipe sections with frequent valve adjustments, or buried pipe sections crossing major traffic arteries. The acquisition process of the high-frequency pressure sequence is as follows: High-sensitivity dynamic pressure sensors are pre-installed at key locations in the target pipe section, such as upstream endpoints, downstream endpoints, or intermediate nodes. These sensors must have a high sampling rate, for example, a sampling frequency of no less than 100Hz, i.e., acquiring at least 100 data points per second, to ensure the capture of details of transient pressure fluctuations caused by water hammer effects and pump start-up and shutdown. The sensors are connected to a data acquisition terminal via an industrial fieldbus or wireless IoT module to record the time-domain changes in fluid pressure inside the pipeline in real time. After analog-to-digital conversion and filtering / denoising, the acquired raw analog signals are processed to form a set of discrete digital signals arranged in chronological order, which constitutes the high-frequency pressure sequence described in this step. .

[0065] like Figure 3 The diagram shown illustrates the vector physical attribute labeling process based on the geometric chord intercept method.

[0066] Reference Figure 3 The figure visually illustrates the effects of steps S1011 and S1012 on the pressure-time two-dimensional projection plane, where the vertical axis represents pressure. The horizontal axis represents time. The horizontal dashed line in the diagram represents the hydrostatic equilibrium baseline. This baseline is a physical reference axis established by identifying steady-state operating time windows and calculating the arithmetic mean, used to eliminate static background pressures such as gravitational potential energy.

[0067] In the figure, the directed line segment connecting adjacent fluctuation amplitude points along the pressure waveform curve represents the pressure evolution chord vector. Based on the comparison between its slope and the pipe's inherent elastic-viscosity critical threshold, two types of vector clusters are marked in geometric space: transient elastic impact vector and steady-state viscous creep vector. The transient elastic impact vector is shown by the dashed arrow at the steep rising edge of the wave crest on the left side of the figure (labeled as "transient elastic impact vector" and "slope too large -> marked as noise"), corresponding to the absolute value of the slope in step S1012. In this situation, the pressure changes extremely rapidly, and the pipe exhibits rigid characteristics similar to Hooke's elasticity; the steady-state viscous creep vector: indicated by the solid arrows in the gently sloping regions of the diagram (labeled as "steady-state viscous creep vector" and "gentle slope -> marked as valid signal"), corresponds to the absolute value of the slope in step S1012. In this case, the pressure change is gradual, and the pipe exhibits irreversible plastic flow characteristics over time.

[0068] Step S1011: Establish a pressure-time two-dimensional projection plane, take the hydrostatic pressure balance baseline as the physical reference axis, and map the discrete pressure sampling points as a set of fluctuation amplitude points relative to the hydrostatic pressure balance baseline.

[0069] In time For horizontal axis, pressure To establish a two-dimensional Cartesian coordinate system with the vertical axis, and to determine the physical reference axis, it is necessary to identify the non-transient operating conditions of the pipe and select a steady-state operating time window. The non-transient operating conditions refer to a state in which the fluid transport system containing the pipe is in a condition free from sudden events such as abrupt valve opening and closing, pump start-up and shutdown, or severe external impacts; that is, the fluid velocity and pressure are in a relatively stable transport state without violent fluctuations. The specific steps for determining the steady-state operating time window are as follows: Set a sliding detection window length. For example, moving the window within the continuously acquired pressure data stream from 10 to 60 seconds; calculating the standard deviation of the pressure data within the current window. and range The range is the difference between the maximum and minimum values; a steady-state determination threshold is set, including the standard deviation threshold. and range threshold If the pressure data within a certain period of time simultaneously meet the following conditions: and If the time period is determined to be the steady-state operating time window, then the arithmetic mean of all discrete pressure sampling points within that window is calculated, and this fixed value is defined as the hydrostatic pressure balance baseline. This baseline serves as a physical reference axis, representing the basic hydrostatic pressure level of the pipe after excluding dynamic disturbances. A mapping operation is performed on discrete points, traversing all discrete pressure sampling points. ,in, Indicates the sequence number of the sampling point. Indicates the first The timestamp of each sampling point Indicates the first The absolute pressure value measured at the sampling point. For each sampling point, a DC component removal calculation is performed, i.e., using the first sampling point... Measured pressure value at each moment Subtract the hydrostatic balance baseline The calculation formula is: Through this calculation, we obtain the first... The fluctuation amplitude relative to the baseline at each moment The original absolute pressure sampling point set is mapped to a fluctuation amplitude point set with the baseline as the zero point. This step aims to eliminate the interference of static background pressure (such as static pressure generated by gravitational potential energy) on dynamic deformation analysis, shifting the analytical perspective to the level of pure pressure fluctuations.

[0070] Step S1011 establishes a two-dimensional pressure-time projection plane and uses the hydrostatic pressure balance baseline as the physical reference axis because the static water pressure in the water supply network is enormous and constant. This pressure, mainly composed of gravitational potential energy and pump station foundation head, can mask subtle dynamic fluctuations caused by pipe creep or minor leaks. By identifying the steady-state window under non-transient conditions, calculating the arithmetic mean, and establishing the baseline, and performing DC component removal, the analysis perspective can be shifted from absolute pressure values ​​to pure fluctuation amplitude levels. This step is indispensable in the overall technical solution because it eliminates the interference of static background factors such as gravitational potential energy on dynamic deformation analysis and establishes the zero-point baseline for all subsequent viscoelastic analyses. Without this baseline, the subsequent quantification of minute deformation increments will lack physical reference coordinates, making it impossible to distinguish between pipe shrinkage and background pressure fluctuations.

[0071] Step S1012: Execute the dynamic geometric chord truncation method, connect adjacent fluctuation amplitude point sets to form pressure evolution chord vectors, and mark the evolution chord vectors in geometric space as transient elastic impact vector clusters and steady-state viscous creep vector clusters based on the slope of the pressure evolution chord vectors and the inherent elastic-viscosity critical threshold of the pipe.

[0072] Connect two adjacent fluctuation amplitude points sequentially according to the time series. and Construct a directed line segment, i.e., a pressure evolution chord vector. Calculate the slope of each chord vector. The inherent elastic-viscosity critical threshold of the pipe material. These physical parameters are obtained through dynamic thermomechanical analysis of standard pipe samples. Specifically, they are based on the characteristic curve of the ratio of the material's storage modulus to its loss modulus as a function of frequency. The ratio of storage modulus to loss modulus is the loss factor. The first derivative of the curve is calculated, i.e., the first derivative of the loss factor with frequency. The frequency corresponding to the minimum negative value of the derivative is determined as the critical frequency point. The minimum negative value represents the largest absolute negative value, i.e., the steepest slope of the curve. This frequency point numerically defines the turning point when polymer chain segments transition from a relaxed response to a rigid response. That is, the frequency of the external force exceeds the characteristic time of chain segment rearrangement, making the material unable to dissipate energy internally and exhibiting a rigid response. This critical frequency is then converted into the corresponding pressure change rate. The calculated absolute value of the slope is then... With threshold Comparison: If This indicates that the pressure changes extremely rapidly, and the pipe exhibits instantaneous elastic deformation. Because the rate of external pressure loading is much higher than the relaxation rate of the material's molecular chains, the polymer chain segments do not have time to untangle and slip. They can only resist external forces through instantaneous reversible changes in molecular bond lengths and bond angles, thus exhibiting rigid characteristics similar to Hooke's elasticity. Therefore, this vector is labeled as a transient elastic impact vector cluster. This indicates that the pressure change is gradual and the pipe exhibits viscous creep. This is because the external pressure loading rate is lower than or close to the relaxation rate of the material's molecular chains, giving the polymer chain segments sufficient time for conformational adjustment, deentanglement, and relative slip between molecules. As a result, it exhibits irreversible plastic flow characteristics that occur over time. Therefore, this vector is labeled as the steady-state viscous creep vector cluster.

[0073] Step S1012 utilizes the dynamic geometric truncated method to construct pressure evolution chord vectors and classifies them based on their slopes and the inherent elastic-viscosity critical threshold of the pipe material. This is because polymer pipes exhibit fundamentally different responses to external forces at different frequencies. By connecting adjacent fluctuation amplitude points to form vectors and calculating their slopes, the rate of pressure change can be intuitively reflected. The critical threshold defines the physical boundary where the material's molecular chain segments transition from a relaxed response to a rigid response. Vectors with large slopes are labeled as transient elastic impact vector clusters, representing instantaneous rigid deformations such as water hammer. Vectors with small slopes are labeled as steady-state viscous creep vector clusters, representing plastic flow caused by molecular chain slippage. This classification is indispensable in the entire method because only by accurately identifying the portion of the pressure history belonging to viscous creep can the physical root cause of false alarms during nighttime events—the hysteretic relaxation effect of the pipe material—be pinpointed, providing a precise data subset for subsequent elimination of this material-related interference.

[0074] like Figure 4 The figure shows a schematic diagram of frustoselective projection based on physical partition marking.

[0075] Reference Figure 4This figure illustrates the viscosity-dominated memory decay cone constructed on a pressure-time two-dimensional projection plane and its selection mechanism for the pressure evolution string vector. The right vertex in the figure represents the current time step. The triangular region that expands to the left in the opposite direction of the time axis is the memory decay cone, and its semi-apex angle is determined by the inherent elastic-viscosity critical threshold of the pipe.

[0076] The diagram visually illustrates the processing logic of two different physical property vectors under the view cone constraint: Transient elastic impact vector (dashed arrow outside the view cone in the diagram): This vector has a large slope, exceeds the boundary of the view cone, and is marked as orthogonally culled / unprojected, representing high-frequency noise or elastic waves that can recover instantaneously. The system ignores its energy contribution. Steady-state viscous creep vector (solid arrow inside the view cone in the diagram): This vector has a gentler slope, falls inside the view cone region, and is marked as an effective component, representing plastic flow that causes hysteresis. The system retains its projected component on the view cone axis and performs weighted aggregation to construct the pipe deformation trajectory history vector.

[0077] Step S1013: On the pressure-time two-dimensional projection plane, with the positive time axis as the central axis and the pipe's inherent elasticity-viscosity critical threshold as the semi-vertical angle, construct an open triangular region in the reverse time axis direction to obtain the viscosity-dominated memory decay cone.

[0078] Using the viscous dominant memory decay cone as a geometric constraint, orthogonal elimination or high-damping attenuation is performed on the transient elastic impact vector cluster, while retaining and enhancing the projection components of the steady-state viscous creep vector cluster on the cone axis. Weighted aggregation is performed according to the time span to construct the pressure evolution string feature synthesis vector.

[0079] The initial elastic modulus of the pipe material is obtained and its reciprocal is calculated to obtain the instantaneous compliance coefficient. The instantaneous compliance coefficient is used as a mapping factor to perform a scalar product transformation on the pressure evolution chord feature synthesis vector to obtain the pipe deformation trajectory history vector.

[0080] On the pressure-time two-dimensional projection plane, at the current time Construct a triangular region with the vertex at the top and expanding along the counter-time axis, pointing towards historical moments, to obtain a memory decay cone dominated by viscosity. The semi-vertex angle of this cone is determined by the arctangent function of the pipe's intrinsic elasticity versus its viscous critical threshold. Based on this view cone as a geometric selection constraint, the pressure evolution chord vector at each historical moment is examined. The tilt angle. For transient elastic impact vector clusters falling outside the view cone, the system determines them to be high-frequency noise or instantaneously recoverable elastic waves. Orthogonal rejection or high-damping attenuation is performed on such vectors. The slope outside the view cone is too large. Orthogonal rejection means mathematically projecting the vector into a direction perpendicular to the view cone boundary and forcing it to zero, or directly setting its amplitude to zero, meaning completely ignoring the energy contribution in that direction. High-damping attenuation means multiplying the vector by a very small attenuation coefficient, such as 0.01. The value of this coefficient is based on the ratio of the material's high-frequency loss modulus to its storage modulus, or determined by experimentally measured instantaneous elastic recovery rate. This assumes that 99% of the instantaneous impact energy is dissipated or recovered in a very short time, without generating long-term memory. This is done to simulate the physical process of high-frequency energy being rapidly dissipated within viscoelastic materials without considering long-term cumulative effects. For steady-state viscous creep vector clusters falling inside the view cone, the system retains their vector magnitude and introduces a time-weighted function based on the Boltzmann superposition principle. It was processed, among which It is a natural constant. Represents the current calculation time. Represents the moment when historical stress occurred. This indicates the time span between the historical event and the present. The characteristic relaxation time constant of the material is assigned by this function as it approaches the current time. The larger the vector, the higher its weight, accurately reflecting the fading characteristic of material memory over time. All effective components after the above filtering and weighting processes are then vector-superimposed. The filtering process removes vectors outside the view frustum, and the weighting process multiplies vectors inside the view frustum by a factor of 1. The superposition here refers to accumulating all the retained, time-weighted vectors according to the parallelogram law or component summation method, and the resulting composite vector is the pressure evolution string feature composite vector. The initial elastic modulus of the pipe material is obtained and its reciprocal is calculated to obtain the instantaneous compliance coefficient. This instantaneous compliance coefficient is used as a mapping factor to perform a scalar product transformation on the pressure evolution string feature composite vector, obtaining the pipe deformation trajectory history vector. Specifically, using the instantaneous compliance coefficient of the pipe material as a mapping factor, a scalar product transformation is performed on the pressure evolution string feature composite vector, mapping the pressure evolution string feature composite vector, which represents the history of pressure fluctuations, from the mechanical load space to the geometric deformation space. The instantaneous compliance coefficient is obtained by performing a standard uniaxial tensile test on the pipe material to obtain its initial elastic modulus, and then calculating its reciprocal. It characterizes the material's ability to undergo elastic deformation at the instant of stress. This transformation process uses this coefficient to numerically convert the cumulative effect of pressure into an equivalent geometric deformation increment, obtaining the pipe deformation trajectory history vector. The synthesized vector essentially represents the equivalent residual deformation trend of the pipe at the current moment caused by the viscous effect accumulated throughout the entire history. It is not only a geometric quantity, but also a physical state quantity. Its core function is that it has the mathematical property of performing tensor product operation with the relaxation spectral density matrix, and can be used as an input source to accurately generate the full history stress-strain memory field tensor in subsequent steps.

[0081] Step S1013, constructing the viscosity-dominated memory decay cone and generating the pipe deformation trajectory history vector, is based on the fact that the material's memory effect decays over time and is primarily dominated by the viscous component. By constructing the cone to orthogonally eliminate or attenuate transient elastic impacts, and using the Boltzmann superposition principle to time-weight the steady-state viscous creep vector within the cone, the physical fact that material memory gradually fades over time can be simulated. Using the pipe's instantaneous compliance coefficient as a mapping factor, a scalar product transformation is performed on the pressure evolution chord feature synthesis vector, converting pressure fluctuations in mechanical space into deformation increments in geometric space. This process is indispensable because it restores abstract pressure data to the actual physical deformation history of the pipe wall. Only by obtaining this deformation trajectory can the equivalent pressure released by the pipe's own deformation be calculated, thus explaining why pressure fluctuations occur in the absence of leakage.

[0082] Step S102: Establish the time-temperature equivalent equation and calculate the shift factor;

[0083] By using a shift factor, the high-frequency pressure sequence on the physical time axis is mapped to the reduced time axis to obtain the de-temperature equivalent pressure sequence;

[0084] Based on the inversion of the material relaxation time spectrum using the de-temperature equivalent pressure sequence, the modulus intensity of each material relaxation time spectrum is arranged according to the relaxation time distribution to construct the material relaxation spectral density matrix.

[0085] Considering that pipes are typical viscoelastic materials, such as high-density polyethylene (HDPE) or polyvinyl chloride (PVC), their mechanical response depends on both time and temperature. To eliminate the interference of ambient temperature fluctuations on the pressure response analysis, a time-temperature equivalent equation is established. Specifically, a standard reference temperature is selected. For example, at 20°C, the shift factor can be calculated using the Williams-Randall-Ferry equation or the Arrhenius equation. The formula is ,in: The actual pipe wall temperature at the time of pressure data acquisition. It is a logarithmic function. and These are empirical constants related to the pipe material. These two constants are typically obtained through dynamic thermomechanical analysis experiments on standard specimens of the same material, or determined by consulting relevant standard property databases of polymer materials. Using this shift factor, the physical time axis will be... High-frequency pressure sequence acquired above Mapped to reduced time axis The above calculation formula is: ,in Represents the current physical observation moment. Let be the integral variable, representing the time from time 0 to the current time. At any point in history between, Representing a historical moment The instantaneous temperature at time t is used to obtain the equivalent pressure sequence after eliminating the influence of temperature history. Based on the generalized Maxwell model, the relaxation time spectrum of the material is inverted using this equivalent pressure sequence. Specifically, based on linear viscoelastic theory, the stress relaxation response and relaxation modulus of the pipe under constant strain are established. There is a direct proportional relationship between them. The equivalent pressure sequence is regarded as a known relaxation modulus evolution sequence, and a series of discrete relaxation time points are set. The relaxation modulus function is fitted using the nonlinear least squares method. ,in, This represents the relaxation modulus as it changes over time. This represents the equilibrium modulus, which is the remaining long-term stiffness of the material as time approaches infinity. This value is obtained as a constant parameter in the fitting process. This represents the total number of Maxwell units connected in parallel in the generalized Maxwell model, i.e., the number of terms in the Proni series; For the summation index, representing the first index. One Maxwell unit Representing the The modulus strength of a relaxed unit. Representing the The characteristic relaxation time points of each relaxation unit Represented by the natural constant The exponential function with base is used for calculation. The resulting modulus strength at each relaxation time point is then calculated. The material relaxation spectral density matrix is ​​constructed in diagonal matrix form based on the relaxation time distribution. The diagonal elements of this matrix That is, the corresponding relaxation time. Relaxed spectral density value This represents the stiffness contribution component or viscoelastic response weight of a material at a specific time scale, revealing the energy distribution of molecular chain segment motion; if a certain corresponding A higher value indicates that this timescale is the dominant stage of material stress relaxation, and a large amount of energy dissipation and molecular chain segment adjustment within the material occurs around this time point.

[0086] Step S103: Perform time-series difference processing on the pipe deformation trajectory history vector to obtain the deformation increment sequence, and perform tensor product operation on the deformation increment sequence and the relaxation spectral density matrix to generate the full history stress-strain memory field tensor.

[0087] Historical vector of pipe deformation trajectory over time Perform a first-order difference operation to generate a deformation increment vector representing the rate of deformation change. Perform tensor product operation to convert the deformation increment vector The relaxation spectral density matrix with relaxation spectral dimension Orthogonal mapping is performed to generate the full-history stress-strain memory field tensor of the two-dimensional structure. The tensor It is a time-frequency memory matrix containing complete thermodynamic state information, and its internal elements The rows correspond to the time dimension, and the columns correspond to the relaxation spectrum dimension, representing the time dimension in the first... The deformation increment occurring at a historical moment, through the first dimension of the relaxation spectrum... The stress contribution components, after decaying through each relaxation time channel and remaining to the current moment, serve as a bridge connecting the Boltzmann superposition principle and deep learning. Specifically, the tensor product operation is performed to generate the full-history stress-strain memory field tensor of the two-dimensional structure. The steps include: With time as the vertical axis, the pipe deformation trajectory history vector corresponds to one row; with the relaxation spectrum as the horizontal axis, corresponding to N discrete relaxation units, and using the relaxation spectral density matrix representing the material's viscoelastic characteristics, each relaxation time channel corresponds to one column. Perform orthogonal mapping operations; for each intersection node in the tensor matrix, i.e., the... Line number The elements of the column are calculated to be the first element. The deformation increment at the first historical moment and the first Modulus coefficient of each relaxed channel And the exponential decay factor calculated based on the time interval from that moment to the current moment. The product of these three factors, through this operation, expands the one-dimensional deformation time series into a two-dimensional time-frequency matrix, which is the full-history stress-strain memory field tensor. It fully records how deformation events at every moment in history are transformed into stress responses, and after decaying through relaxation channels of different speeds, they make a physical contribution to the pressure at the current moment, thus providing a deterministic input containing complete thermodynamic state information for subsequent neural networks.

[0088] Step S103, which involves differencing the pipe deformation trajectory history vector to obtain the deformation increment sequence and performing a tensor product operation with the relaxation spectral density matrix, is based on the fact that stress relaxation in polymer materials is not a decay of a single time constant, but rather a multiple relaxation process covering a broad time domain. Through tensor product operations, the deformation increment in the time dimension is orthogonally mapped to the modulus distribution in the relaxation spectral dimension, generating a full-history stress-strain memory field tensor. This tensor completely records how deformation events at each historical moment decay in different relaxation channels and remain until the current moment. This step is indispensable because it constructs a time-frequency matrix containing complete thermodynamic state information, providing deterministic physical input for the subsequent neural network, enabling it to understand how complex historical deformations accumulate and lead to spurious pressure fluctuations at the current moment.

[0089] Step S200: Construct a neural hysteresis operator network, take the full history stress-strain memory field tensor as input, perform nonlocal memory characteristics and hysteresis loop analysis, and obtain the intrinsic nonlinear hysteresis response signal of the pipe.

[0090] The described neural hysteresis operator network is a hybrid architecture deeply coupled with a physical hysteresis model. Its construction process first defines an input layer that receives a full-history stress-strain memory field tensor containing full-dimensional time-frequency information. This tensor then enters a multi-head attention mechanism layer, where linear transformations and scaling dot product operations on query vectors, key vectors, and value vectors are used to calculate the influence weights of historical moments on the current state and generate a weighted fusion feature vector. Next, a state mapping layer is constructed, using matrix multiplication to interact the weighted fusion feature vector with the original input tensor to generate an equivalent relaxed state vector. This relaxed state vector is then compressed and mapped into a scalar equivalent driving force using a set of trainable linear weights and biases. The core layer of the network consists of several parallel layers. The network consists of a set of PI neurons, each of which is a discrete microscopic hysteresis unit. It encapsulates a weighted stopping operator recursive logic based on a specific yield threshold to simulate the elastic retention and plastic slip behavior of materials under different driving torque values. The output layer of the network linearly weights and superimposes the internal states of all PI neurons to generate the intrinsic nonlinear hysteresis response signal of the pipe. The entire network is then subjected to backpropagation algorithm, with measured delinearized hysteresis deformation data as the supervision target. The attention weights, mapping parameters, and threshold distribution and weight coefficients of PI neurons are jointly iteratively optimized to establish a mathematical model that can accurately characterize the nonlocal memory characteristics and hysteresis loop morphology of the target pipe.

[0091] Step S201: The network input layer receives the full history stress-strain memory field tensor, and uses a multi-head attention mechanism to weight the influence of different historical moments on the current state to obtain a weighted fusion feature vector;

[0092] The full-history stress-strain memory field tensor It refers to the construction that includes the time dimension. In neural network operations, the relaxation spectrum dimension of a two-dimensional matrix serves as the feature dimension of the network, denoted as [missing information]. , This represents the number of discrete historical moments contained in the full-history stress-strain memory field tensor along the time dimension. Features are extracted using a multi-head attention mechanism; specifically, the number of attention heads is set to h, for example, h=8. For the ... Each attention head maps the input tensor to a query vector through a linear transformation. Key vector Sum value vector Using the scaling dot product attention formula Perform calculations, its Query features representing the current state Index key value representing a historical moment, Represents the transpose of the index key value at a historical moment. Bearing specific residual stress values, The dimension of the key vector. This represents the attention score matrix, where Softmax is a normalized exponential function, and each element represents the model's judgment at a certain past moment. The residual stress state at the current moment Importance weights for material deformation prediction, score matrix and value vector Multiplication achieves a weighted summation of all historical residual stress information, thus combining all... The outputs of each feature vector are concatenated and fused through a linear transformation to produce a weighted fused feature vector.

[0093] Step S202: Perform matrix multiplication on the weighted fused feature vector and the full-history stress-strain memory field tensor to obtain the equivalent relaxed state vector;

[0094] Map the equivalent relaxed state vector space to a scalar equivalent driving force;

[0095] Based on the scalar equivalent driving force, parallel PI neurons are used, each neuron being a stopping operator. The threshold distribution function of the operator is optimized through a deep learning backpropagation algorithm to fit the nonlinear hysteresis of the pipe and obtain the intrinsic nonlinear hysteresis response signal of the pipe.

[0096] Weighted fusion feature vector for A time-weighted aggregation of a column vector over a tensor M: The column vector... Transpose row vectors ,and Full-history stress-strain memory field tensor Perform matrix multiplication to obtain Equivalent relaxed state vector of dimension ,Right now ,in Indicates that at the current moment, along the first... The equivalent residual stress or equivalent residual strain intensity accumulated in each relaxed modal channel; perform spectral domain mapping projection to transform the equivalent relaxed state vector. from The dimensional relaxation mode space is mapped to a scalar equivalent driving force. In a neural network, a set of trainable weight parameter vectors is set. and bias parameters ,in for 3D column vector, For scalar constants, through a linear mapping relationship Calculate the scalar equivalent driving force at the current moment. The This step serves as the unified input for subsequent hysteresis operators at the micro level. The complex relaxation states of each modality are comprehensively mapped to a single scalar signal driving the hysteresis behavior of the material at the macroscopic level. A parallel PI neuron set is constructed, which is a neural network implementation of a large number of infinitesimal stopping operators in the classic Preisach-Ishlinskii hysteresis model. Each PI neuron corresponds to a discretized microscopic hysteresis unit, which can be activated by the input driving force. The repeated loading-unloading process generates a hysteresis response with memory characteristics. Specifically, let the number of the parallel PI neurons be... For index is ( ) Each PI neuron is pre-assigned a trainable yield threshold. and a trainable weight coefficient The yield threshold The weighting coefficients represent the minimum driving force amplitude required for the microscopic hysteresis unit to transition from an elastic response state to a plastic or irreversible response state. This characterizes the contribution of the micro-unit to the global hysteresis loop; during the forward computation of the neural network, for each discrete time step... ( ), The maximum value of the discrete time step is represented by scalar equivalent driving force at the current moment. Maintain internal state variables for each PI neuron. This internal state is used to record the equivalent hysteresis deformation formed by the microscopic hysteresis unit under each loading history, and its numerical update rule adopts a recursive relationship in the form of a weighted stopping operator:

[0097] ;

[0098] in, This indicates taking the maximum value. This indicates taking the minimum value. The internal state value at the previous moment. The physical meaning of the above update rule, which represents the yield threshold of this unit, is that when the current equivalent driving force... In the range Within this range, the neuron's output remains within the boundary of that range and exhibits stopping properties, thus simulating the elastic memory behavior of a material when its yield threshold is not exceeded; when When outside the above range, Bounded updates occurring in directions beyond indicate that the material undergoes irreversible yielding or plastic slip in that micromode, exhibiting the opening characteristics of the hysteresis loop.

[0099] At each time step By linearly superimposing the internal states of all PI neurons, the intrinsic nonlinear hysteresis response signal of the pipe is obtained. , For the current moment The intrinsic nonlinear hysteresis response signal of the pipe generated by the parallel weighted stopping operator network is shown below. As the response signal changes repeatedly over time, it can form a closed or semi-closed nonlinear hysteresis curve on the scalar equivalent driving force-equivalent hysteresis deformation plane; during the network training phase, the intrinsic nonlinear hysteresis response signal of the pipe obtained by the above-mentioned parallel PI neurons is used. The measured total deformation response data is compared with the actual scalar equivalent driving force-total deformation response data obtained from standard loading-unloading tests or long-term operational monitoring of the pipe material. The measured total deformation response data is then delinearized: a linear elastic response component of the pipe material is fitted using the least squares method or endpoint connection method, and this linear elastic response component is subtracted from the measured total deformation response data to extract the hysteresis deformation sequence. To eliminate dimensional differences and unify the data scale, the hysteresis deformation sequence is normalized, for example, by max-min normalization, to obtain the... Normalized target reference value at each moment ,right Equal normalization processing to ensure and Under the same normalized dimension, the intrinsic nonlinear hysteresis response signal of the pipe material calculated by the above parallel PI neurons is... Compared with the target reference value Compare and construct the loss function. For example, using the mean square error form The parameters were analyzed using the backpropagation algorithm. , and the yield threshold of each PI neuron and weighting coefficients Gradient updates are performed, and these parameters are iteratively adjusted using optimization algorithms such as Adam or SGD, so that the hysteresis loop output by the neural network gradually approximates the nonlinear hysteresis loop of the real pipe in terms of shape, area, and yield inflection point location; after the above training process is completed, a set of parameters is obtained. , , and That is, a set of threshold distribution functions and weight distributions that match the material properties of the target pipe were determined.

[0100] Step S200, constructing a neural hysteresis operator network to obtain the intrinsic nonlinear hysteresis response signal of the pipe, is based on the fact that the viscoelastic response of the pipe is highly nonlinear and has strong hysteresis, making it difficult to accurately describe using traditional linear equations. Features are extracted through a multi-head attention mechanism and mapped to scalar equivalent driving forces, which then drive a parallel PI neuron set. The memory properties of its stopping operator are utilized to fit the nonlinear hysteresis loop of the pipe. This step is indispensable because it acts as a digital twin model, capable of accurately predicting the intrinsic nonlinear hysteresis response signal of the pipe at the current moment—the ghost signal that caused the false alarm—based on previous pressure history, providing a minuend for subsequent signal stripping.

[0101] Step S300: Perform point-by-point differential processing on the high-frequency pressure sequence and the intrinsic nonlinear hysteresis response signal of the pipe to remove non-stationary background fluctuations, avoid the modulation interference of material nonlinearity on the fluid signal, and extract the fidelity fluid dynamics anomaly residual.

[0102] Step S301: Match the high-frequency pressure sequence with the intrinsic nonlinear hysteresis response signal of the pipe on the time axis point by point;

[0103] The term "point-to-point correspondence on the timeline" refers to ensuring, through hardware synchronization or software alignment, that at every specific point in time, the acquired high-frequency pressure sequence and the calculated intrinsic nonlinear hysteresis response signal of the pipe belong strictly to the same moment. The system binds these two signal values ​​together, forming a one-to-one synchronous data pair, and stores them sequentially according to their chronological order. The purpose of this is to eliminate time differences caused by signal transmission or calculation, ensuring that subsequent comparisons or operations on these two signals are performed based on exactly the same physical moment, thereby ensuring the accuracy of data processing.

[0104] Step S302: Based on the subtraction operation performed on the subsequent intrinsic nonlinear hysteresis response signal of the pipe, the preliminary residual is obtained, and the residual is decomposed into intrinsic mode functions by Hilbert-Huang transform.

[0105] To eliminate differences in physical dimensions and accurately preserve waveform characteristics, preset system calibration extreme value parameters were used to analyze the time-aligned high-frequency pressure sequences. and the intrinsic nonlinear hysteresis response signal of pipe material The normalization calculation is performed using the following formulas: and ,in This represents the normalized high-frequency pressure sequence. This represents the normalized intrinsic nonlinear hysteresis response signal of the pipe material. It is the calibrated maximum pressure. This involves calibrating the maximum response amplitude and mapping both signals to a dimensionless interval, such as [0,1]. Normalization is performed before subtraction. Physically, this is because the high-frequency pressure sequence and the hysteresis response signal have different physical dimensions, making direct subtraction meaningless. Logically, normalization transforms both signals into relative intensity or percentage changes relative to their respective system limits, thus establishing a unified comparison benchmark. Subtraction is then performed based on this benchmark. The obtained normalized preliminary residuals This effectively isolates the fundamental trends in the signal, directly reflecting the degree of nonlinear distortion and high-frequency abnormal fluctuations in the actual pipe response relative to the input pressure, providing a clean and comparable dimensionless data object for subsequent analysis. The empirical mode decomposition method in the Hilbert-Huang transform is used to analyze the aforementioned preliminary residual signal. Multi-scale decomposition is performed to extract intrinsic mode functions, specifically: identifying the preliminary residual signal. For all local maxima and local minima, the upper envelope of the signal is fitted using a cubic spline interpolation function. and lower envelope Calculate the mean curves of the upper and lower envelopes. The mean curve is then subtracted from the original residual signal to obtain the intermediate component. At this point, it is necessary to make a judgment. Whether the intrinsic mode function (IMF) satisfies the two defining conditions is: first, the number of extreme points is equal to or differs from the number of zero-crossing points by at most 1; second, the mean of the upper and lower envelopes at any point must be zero. If the above conditions are met, then it is confirmed as a first-order eigenmode function. If not satisfied, then As a new signal to be analyzed, the above screening process is repeated until the conditions are met. Through this iterative process, the initial residual signal is... Ultimately decomposed into Eigenmode function components at different frequency scales , For the first The intrinsic mode function of order.

[0106] Step S303: According to the Shannon entropy criterion, the slow-varying characteristic IMF component of the pipe is removed from the intrinsic mode function, and the pure fluid dynamics anomaly residual is reconstructed.

[0107] For the decomposed The intrinsic mode function components are selected sequentially. Each component Perform calculations, where The order index of the intrinsic mode function is given, and its value range is given. to , This represents the total number of IMF components obtained from the decomposition. To accurately measure the degree of information disorder contained in each component, the th... The total energy of each component The calculation formula is: ,in This represents the component at time [time]. The L2 norm, or modulus, physically characterizes the instantaneous amplitude intensity of this component at any given time point. The summation is performed by taking the squares of the amplitudes; based on this total energy, the components are defined at each time step. Normalized probability distribution This distribution represents the proportion of signal energy in the total energy at a given moment. Based on this, the energy Shannon entropy of this component is calculated using the Shannon entropy formula. The calculation formula is: Physically, Shannon entropy... The magnitude of entropy directly reflects the complexity and uncertainty of the signal: a lower entropy value indicates a more regular and ordered signal, corresponding to a deterministic low-frequency trend; a higher entropy value indicates a more chaotic and random signal, corresponding to abrupt signals containing rich transient information. Based on a preset entropy threshold... For each energy Shannon entropy calculated Logical discrimination and filtering are performed to separate the physical source of the signal; this threshold... The baseline was determined by pre-analyzing the entropy distribution characteristics of low-frequency background noise generated by pipe material under normal operating conditions due to creep, relaxation, or thermal expansion and contraction. The specific screening and reconstruction process involves: calculating the... With threshold Compare them one by one, if Then determine the first Individual eigenmode functions This mainly characterizes the slow-changing features of the pipe itself, such as low-frequency, regular fluctuations caused by material relaxation, which are considered background interference and are therefore removed; if Then determine the The components contain high-frequency transient information from fluid dynamics, such as turbulence, pressure waves, or cavitation phenomena caused by leakage. These are valid signals to be detected and are therefore retained. All components meeting the retention criteria are included. The components are linearly superimposed to reconstruct a pure fluid dynamics anomaly residual that has removed the interference from the pipe material. This signal will serve as input data for subsequent high-sensitivity leak detection, improving the signal-to-noise ratio of the detection.

[0108] Steps S301 to S303, which align and differencing the high-frequency pressure sequence with the intrinsic nonlinear hysteresis response signal of the pipe and screen the intrinsic mode functions based on the Shannon entropy criterion, are based on the fact that physical subtraction is the most direct means of eliminating background interference. Through point-by-point alignment and normalized difference along the time axis, the intrinsic fluctuations of the pipe predicted by the neural network are stripped away, yielding a preliminary residual. The Hilbert-Huang transform is used to decompose the residual into intrinsic mode functions at different frequency scales, and the energy Shannon entropy is calculated. The degree of disorder in the signal is judged based on the entropy value, eliminating low-entropy slow-varying residual features of the pipe and retaining high-entropy fluid dynamics abrupt change information. This process is indispensable because it achieves layer-by-layer filtering from time-domain differencing to frequency-domain screening, ensuring that the finally extracted signal is a pure fluid anomaly residual after removing material memory interference.

[0109] Step S304: Extract the operator density distribution features of PI neurons and the multi-scale entropy features of each IMF component to obtain the first feature vector and the second feature vector. Concatenate the first feature vector and the second feature vector to obtain a high-dimensional joint feature vector and construct the hysteresis-entropy conjugate feature mapping space.

[0110] Based on the trained PI neural network, a set of parameters characterizing the memory properties of pipe materials is extracted, that is, for a quantity of... Parallel PI neurons are used to obtain a set of weight coefficients after training convergence. ,in The weighting coefficients of this group Arranged according to the neuron index order The dimensional vector is defined as the operator density distribution feature of the PI neuron, denoted as the first eigenvector. This vector physically maps the hysteresis operator distribution density of the pipe under different stress levels; for each eigenmode function component selected and retained in the previous step... Perform multi-scale entropy analysis on it, that is, set the scale factor. The maximum value is (For example, take 10), calculate the Components in scale factors from 1 to The set of sample entropy values ​​below, arranged in order to form The dimensional vector is defined as the multi-scale entropy feature of each IMF component, denoted as the second feature vector. , Let be the sample entropy value under the maximum scaling factor; for each Components, and their corresponding first feature vectors The second eigenvector unique to this component Perform feature fusion, first feature vector Since material properties are globally shared background attributes across all components, the same set of material features, or a subset optimized for a specific frequency band, is used here. Specifically, two vectors are concatenated to form a vector with dimension [missing information]. High-dimensional joint feature vector The set of joint feature vectors corresponding to all retained IMF components. The generated multidimensional Euclidean space is a hysteresis-entropy conjugate feature map space. Each data point in this space is represented in the conjugate form of material memory and fluid entropy, which fully describes the coupling state of the signal components in the two dimensions of material nonlinearity and fluid complexity.

[0111] Step S305: Perform orthogonal subspace decomposition on the high-dimensional joint feature vector in the hysteresis-entropy conjugate feature mapping space, construct the material memory principal plane and the fluid disturbance orthogonal complementary plane, and perform vector decoupling on the signal components to obtain the material memory component vector and the fluid disturbance component vector;

[0112] Based on the constructed hysteresis-entropy conjugate feature map space, a deep decoupling analysis of the signal is performed. This involves generating a dimension-based feature map for each retained IMF component within this space. High-dimensional joint feature vector As row vectors, they are arranged sequentially according to component index order and stacked to form a distribution state feature matrix. The distribution state feature matrix can characterize the distribution state of the entire signal in the material-fluid coupling dimension. Principal component analysis or singular value decomposition algorithms are applied to this feature matrix. Considering that the creep and relaxation of the pipe material belong to a global, slowly varying background, and exhibit high-energy dominant directions in the feature space, the top eigenvalues ​​are selected. The subspace spanned by the principal components is defined as the material memory principal plane, which primarily captures the data generated by the first eigenvector. The reflected PI operator density distribution information; based on the orthogonal complement principle in linear algebra, a complement space completely orthogonal to the principal plane is constructed, defined as the fluid disturbance orthogonal complement surface, which mainly retains the second eigenvector. The multi-scale entropy abrupt change information characterized by fluid leakage is represented; vector decoupling is performed by calculating inner products or matrix multiplication to decouple each... Projecting the vectors onto the two orthogonal subspaces, respectively, we calculate the projection components of the vectors onto the basis vectors of the material memory principal plane, and reconstruct the material memory component vectors. , Characterize the background noise of the pipe by calculating its projection components onto the orthogonal complementary base vector of the fluid disturbance, and reconstruct the fluid disturbance component vector. , It characterizes pure leakage signals and completely achieves mathematical separation of the nonlinear memory characteristics and fluid dynamic characteristics of pipes at the geometric space level.

[0113] Step S306: Calculate the ratio of the L2 norm of the fluid disturbance component vector to the sum of the L2 norms of the material memory component vector and the fluid disturbance component vector to obtain the fluid dynamics confidence coefficient;

[0114] The pure fluid dynamics anomaly residuals are reconstructed using soft threshold weighting using the fluid dynamics confidence coefficients to synthesize authentic fluid dynamics anomaly residuals.

[0115] Based on material memory component vectors With fluid disturbance component vector The fluid dynamics confidence coefficient is calculated and soft-threshold weighted reconstruction is performed to synthesize high-fidelity fluid dynamics anomaly residuals. Specifically, for each IMF component, the index is... The signal component tendency is quantitatively evaluated in the feature space by calculation. The L2 norm, i.e., the vector magnitude, represents the energy of fluid disturbance; calculation vector magnitude and add The ratio of the total modulus to the total modulus, where the total modulus is the L2 norm, yields the hydrodynamic confidence coefficient. The value ranges from 0 to 1, and this coefficient mathematically measures the probability density that the current component contains true leaked information. A nonlinear soft thresholding model based on the Sigmoid function or exponential function is constructed to assign the confidence coefficient of each component to the given value. Convert into corresponding signal reconstruction weights This results in high-confidence components projected onto the principal direction of fluid disturbance in the feature space receiving a weight close to 1, while the weight of background noise components projected onto the material memory principal plane is adaptively attenuated to close to 0. The principal direction of fluid disturbance is the orthogonal complement of the material memory principal plane. For those components that, although they passed the S303 entropy value screening, were found by S305 analysis to actually contain a large number of nonlinear material memory features, i.e. The components with a large proportion are given extremely low weights for secondary suppression, while components with significant fluid characteristics are given high weights. All weighted IMF components are superimposed in the time domain to output a high-fidelity fluid dynamics anomaly residual. This residual, based on the initial noise reduction of S303, further eliminates deep nonlinear material memory interference, providing a data foundation with an extremely high signal-to-noise ratio for subsequent detection.

[0116] The rationale for constructing the hysteresis entropy conjugate feature mapping space and performing orthogonal subspace decomposition and soft-threshold reconstruction in steps S304 to S306 is that single-dimensional filtering may not be able to completely separate strongly coupled nonlinear signals. By extracting the operator density distribution features of PI neurons and the multi-scale entropy features of IMF components and concatenating them into a high-dimensional joint feature vector, a mapping space containing both material memory and fluid disturbance attributes is constructed. Orthogonal decomposition is performed in this space to decouple the signal vector into material memory components and fluid disturbance components, and confidence coefficients are calculated accordingly for soft-threshold weighting. This step is indispensable; it achieves deep cleaning of residual noise at the geometric feature space level, ensuring that the synthesized high-fidelity fluid dynamics anomaly residual has an extremely high signal-to-noise ratio and completely eliminating the interference of nonlinear material properties.

[0117] Step S400: Reconstruct the fidelity fluid dynamics anomaly residual into a high-dimensional phase space trajectory, construct a spatiotemporal evolution Poincaré cross section, input it into a convolutional neural network, and determine the real leakage negative pressure wave and sensor random drift through the attractor morphology differences on the spatiotemporal evolution Poincaré cross section.

[0118] Step S401: Reconstruct the high-dimensional phase space trajectory from the fidelity fluid dynamics anomaly residuals using Takens' embedding theorem;

[0119] Using a high-fidelity fluid dynamics anomaly residual sequence, denoted as ,in Represents the overall sequence of anomaly residuals in high-fidelity fluid dynamics. Represents a single outlier residual data point in the sequence. This represents the total length of the data points contained in the sequence. Calculate the optimal delay time for this sequence. and optimal embedding dimension Among them, the optimal delay time The optimal embedding dimension is determined by the mutual information method, specifically by calculating the delay time corresponding to the first minimum value of the mutual information function of the sequence. The minimum dimension is determined using the spurious neighbor method, specifically, the dimension at which the proportion of spurious neighbors approaches zero as the dimension increases. Based on this determined value... and Construct phase space vectors based on Takens' embedding theorem. ,in Indicates the first The reconstructed phase space state vector is expressed as follows: ,in This represents the time index number of the phase space vector, with a value range of [value range missing]. , here The total number of phase space vectors generated after reconstruction is represented by the formula: All phase space vectors Connect them in chronological order to form a... trajectory matrix of evolution in dimensional Euclidean space This matrix is ​​the reconstructed high-dimensional phase space trajectory, which can be derived from a one-dimensional time series. The nonlinear dynamic characteristics of the fluid dynamics system are recovered.

[0120] Step S402: Extract the Poincaré section in phase space to transform the continuous dynamic trajectory into a discrete two-dimensional point cloud distribution map;

[0121] In the construction In the high-dimensional phase space, a suitable hyperplane is selected as the Poincaré section. To maximize the capture of dynamic abrupt changes caused by fluid leakage, the plane formed by the two principal component directions with the greatest variability in the phase space trajectory is chosen as the projection reference for the section, or the plane passing through the geometric center of the phase space trajectory and perpendicular to the average flow direction of the trajectory is selected as the Poincaré section. All unidirectional intersection points between the high-dimensional phase space trajectory and the Poincaré section are recorded, i.e., the points where the trajectory passes from one side of the section to the other, resulting in a series of discrete intersection point sets. ,in This represents the total number of intersection points. These high-dimensional intersection points are mapped to a two-dimensional plane coordinate system to generate a two-dimensional scatter plot. To enhance image features, the two-dimensional scatter plot is binarized and normalized, marking the pixel position of the scatter points as 1 (white) and the background as 0 (black), generating a plot with a size of [missing information]. A black and white binary image, in which Represents the height of the image in pixels. Represents the width of the image in pixels. For example, a 224×224 pixel image is a discrete two-dimensional point cloud distribution map. This dimensionality reduction map preserves the periodicity, quasi-periodicity, or chaotic characteristics of the system's motion, intuitively reflecting the stability state of the fluid system.

[0122] Step S403: Input the discrete two-dimensional point cloud distribution map into the convolutional neural network to perform steady-state attractor pattern recognition and chaotic divergence trajectory analysis to determine the real leakage negative pressure wave and sensor random drift; wherein, steady-state attractor pattern recognition includes convergent point or closed loop structure analysis, and chaotic divergence trajectory analysis includes divergent singular attractor or non-closed fracture structure analysis.

[0123] A lightweight convolutional neural network model is constructed. This model architecture includes an input layer for receiving the pixel matrix of the point cloud, several convolutional layers for extracting the spatial distribution features of the point cloud, pooling layers for dimensionality reduction while preserving feature invariance, and fully connected layers. A training dataset is established, and topological morphology labels are defined. Historical data is collected to generate Poincaré cross-sectional point cloud images. Based on the geometric topological features of the point clouds, they are classified into two categories: normal / interference and leakage. The computer quantization criteria for each category are as follows: Normal / interference (label 0) corresponds to a point cloud exhibiting convergent point-like or closed loop-like structures. Convergent point-like structures are defined as discrete points highly clustered on a two-dimensional plane, where the Euclidean distance from all points to a geometric center is less than a preset convergence radius threshold. Specifically, this manifests as the standard deviation of the coordinates of the point cloud set being less than a preset convergence threshold. ,Right now and , and These represent the standard deviations of the x and y coordinates of all points in a discrete point cloud set in a two-dimensional plane coordinate system, respectively. They are used to mathematically quantify the degree of dispersion of the point cloud in the horizontal and vertical directions, with more than 95% of the points falling within a circle centered at the geometric centroid with a radius of [missing information]. Within the covered circle, the system is characterized as being in a stable equilibrium state or dominated by background noise; a closed loop is defined as discrete points arranged sequentially on a plane to form one or more geometrically continuous closed curves, with uniform average adjacent distances between points, specifically represented by the radial distance variance of the point cloud set relative to its geometric centroid. Less than the minimum value Simultaneously, the angular distribution entropy of the points exceeds the threshold, indicating that the system is under periodic vibration or regular pump pulsation disturbance. Leakage type (label 1) corresponds to point clouds exhibiting divergent singular attractors or irregular fracture structures. A singular attractor is defined as a fractal structure with self-similarity exhibited by discrete points. The point cloud distribution occupies a specific phase space region but is locally infinitely nested; specifically, the point cloud set has a non-integer box-counting dimension. ,like Furthermore, if the spatial distribution histogram of the point cloud exhibits a non-Gaussian distribution, it is identified as a strange attractor, representing a chaotic dynamic change caused by a small leakage in the system. Irregular fracture structures are defined as the point cloud's original continuous trajectory being broken, exhibiting a disordered, multi-clustered discrete scattering pattern. Specifically, this is manifested by the number of connected components in the point cloud image being greater than 1. That is, performing connectivity analysis on the binarized point cloud image detects two or more independent, unconnected pixel clusters with an Euclidean distance greater than the clustering threshold, representing a turbulent abrupt change caused by the disruption of system boundary conditions. The CNN model is trained under supervision using the labeled dataset. Forward propagation is used to calculate the prediction results, and the cross-entropy loss function is used to quantify the error between the predicted value and the true mathematical topological label. Backpropagation is used to calculate the gradient, combined with an optimizer to iteratively update the network weights, enabling the model to learn nonlinear mapping weights from the image pixel space to different topological forms. Finally, in the real-time detection stage, the discrete two-dimensional point cloud distribution map is input into the trained CNN model. The model's output layer calculates the confidence probability of the input image belonging to each category using the Softmax function. The decision logic is: if the model outputs the probability of belonging to the leakage class... If the value is greater than or equal to a preset alarm threshold, such as 0.85, then a real leak event is determined to have occurred within the current fluid system, meaning the model mathematically confirms that the image features conform to the distribution pattern of singular attractors or fracture structures; conversely, if the value is less than or equal to the preset alarm threshold, then a real leak event has occurred within the fluid system. If the noise level is less than the threshold, it is determined to be background noise or periodic interference, and the system has no leakage. Through this step, the topological features of the Poincaré cross section are automatically quantified using deep learning, and the nonlinear dynamic mutations caused by tiny leakage are accurately identified.

[0124] Step S400, which reconstructs the high-dimensional phase space trajectory from the fidelity fluid dynamics anomaly residuals and uses a convolutional neural network to identify the Poincaré cross-section attractor morphology, is based on the fact that the negative pressure wave generated by fluid leakage and sensor drift differ in their topological structure due to their inherent dynamic nature. By reconstructing the phase space using Takens' embedding theorem and extracting the Poincaré cross-section, a one-dimensional time series can be transformed into a two-dimensional point cloud distribution map revealing the system's motion. Real leakage manifests as divergent singular attractors or fractured structures, while drift manifests as convergent point-like or closed-loop structures. This step is indispensable because it transcends traditional time-domain amplitude judgment, making the final distinction from the perspective of nonlinear dynamic topological morphology, effectively differentiating real leakage from random drift, and completely solving the problem of periodic false alarms.

[0125] Example 2

[0126] like Figure 2 As shown, an intelligent leak detection device for underground pipelines includes the following:

[0127] The memory conversion module is used to acquire the high-frequency pressure sequence of the target pipe section. Based on the Boltzmann superposition principle, the high-frequency pressure sequence is converted into the full-history stress-strain memory field tensor by using the ambient temperature as a thermorheological adjustment factor.

[0128] The hysteresis response module is used to construct a neural hysteresis operator network, taking the full-history stress-strain memory field tensor as input, performing nonlocal memory characteristics and hysteresis loop analysis to obtain the intrinsic nonlinear hysteresis response signal of the pipe.

[0129] The abnormal residual module is used to perform point-by-point differential processing on the high-frequency pressure sequence and the intrinsic nonlinear hysteresis response signal of the pipe, remove non-stationary background fluctuations, avoid the modulation interference of material nonlinearity on the fluid signal, and extract the fidelity fluid dynamic abnormal residual.

[0130] The leakage determination module is used to reconstruct the fidelity fluid dynamics anomaly residual into a high-dimensional phase space trajectory, construct a spatiotemporal evolution Poincaré cross section, input it into a convolutional neural network, and determine the real leakage negative pressure wave and sensor random drift through the attractor morphology differences on the spatiotemporal evolution Poincaré cross section.

[0131] It should be noted that the hyphen "-" in the text only indicates the relationship or correspondence between two physical quantities.

[0132] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0133] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0134] The foregoing description is illustrative of the invention and should not be construed as limiting it. Although several exemplary embodiments of the invention have been described, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the invention. Therefore, all such modifications are intended to be included within the scope of the invention as defined in the claims. It should be understood that the foregoing description is illustrative of the invention and should not be construed as limiting it to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The invention is defined by the claims and their equivalents.

Claims

1. A method of intelligent leak detection for underground pipes, characterized in that, The method comprises the following steps: Collecting a high-frequency pressure sequence of a target pipe section, establishing a time-temperature equivalence equation to calculate a shift factor based on the viscoelastic rheological response law of the pipe material, and mapping the high-frequency pressure sequence on the physical time axis to a reduced time axis using the shift factor to obtain a temperature-free equivalent pressure sequence; and converting the temperature-free equivalent pressure sequence into a full historical stress-strain memory field tensor; Constructing a neural hysteresis operator network, taking the full historical stress-strain memory field tensor as input, and performing non-local memory characteristic and hysteresis loop analysis to obtain an intrinsic nonlinear hysteresis response signal of the pipe material; Performing point-by-point difference processing on the high-frequency pressure sequence and the intrinsic nonlinear hysteresis response signal of the pipe material to strip non-stationary background fluctuations, avoid the modulation interference of material nonlinearity on fluid signals, and extract a faithful fluid dynamics anomaly residual; Reconstructing the faithful fluid dynamics anomaly residual into a high-dimensional phase space trajectory, constructing a spatiotemporal evolution Poincare section, inputting it into a convolutional neural network, and determining a real leakage negative pressure wave and sensor random drift through the difference in attractor morphology on the spatiotemporal evolution Poincare section; The analysis step of the full historical stress-strain memory field tensor comprises: Collecting a high-frequency pressure sequence of a target pipe section, using a viewing cone to filter and time-weighting processing stress components based on the viscoelastic rheological response law of the pipe material, and constructing a pipe material deformation trajectory history vector; Establishing a time-temperature equivalence equation to calculate a shift factor; Mapping the high-frequency pressure sequence on the physical time axis to a reduced time axis using the shift factor to obtain a temperature-free equivalent pressure sequence, and determining the reduced time axis by the shift factor; Based on the temperature-free equivalent pressure sequence, inverting the material relaxation time spectrum, arranging the modulus strength of the material relaxation time spectrum according to the relaxation time distribution to construct a material relaxation spectrum density matrix; Performing time series difference processing on the pipe material deformation trajectory history vector to obtain a deformation increment sequence, and performing tensor product operation on the deformation increment sequence and the relaxation spectrum density matrix to generate a full historical stress-strain memory field tensor; The analysis step of constructing the pipe material deformation trajectory history vector comprises: Establishing a pressure-time two-dimensional projection plane, taking the fluid static pressure balance reference line as a physical reference axis, and mapping discrete pressure sampling points to fluctuation amplitude point sets relative to the fluid static pressure balance reference line; Performing dynamic geometric chord intersection, connecting adjacent fluctuation amplitude point sets to form pressure evolution chord vectors, and marking the evolution chord vectors as transient elastic impact vector clusters and steady-state viscous creep vector clusters in geometric space according to the slope of the pressure evolution chord vectors and the inherent elastic-viscous critical threshold of the pipe material; The analysis step of constructing the pipe material deformation trajectory history vector further comprises: On the pressure-time two-dimensional projection plane, taking the positive direction of the time axis as the central axis and the inherent elastic-viscous critical threshold of the pipe material as the half-apex angle to construct an inverse time axis direction opening triangle region to obtain a viscous dominant memory decay viewing cone; With the viscoelastic memory cone as a geometric constraint, orthogonal culling or high-damping attenuation is performed on the transient elastic impact vector cluster, the projection component of the steady-state viscous creep vector cluster on the cone axis is retained and enhanced, and the pressure evolution string characteristic synthesis vector is constructed by weighted aggregation according to the time span; The initial elastic modulus of the pipe material is obtained, and the reciprocal thereof is calculated to obtain the instantaneous compliance coefficient; the pressure evolution string characteristic synthesis vector is subjected to scalar multiplication transformation by using the instantaneous compliance coefficient as a mapping factor to obtain a pipe deformation trajectory history vector; The steps of the non-local memory characteristic and the hysteresis loop analysis include: The network input layer receives a full history stress-strain memory field tensor, and the influence weight of different historical moments on the current state is obtained by a multi-head attention mechanism to obtain a weighted fusion feature vector; The weighted fusion feature vector is subjected to matrix multiplication operation with the full history stress-strain memory field tensor to obtain an equivalent relaxation state vector; The equivalent relaxation state vector is spatially mapped into a scalar equivalent driving force; Based on the scalar equivalent driving force, a parallel P-I neuron is used, each neuron is a stop operator, and the threshold distribution function of the operator is optimized by a deep learning back propagation algorithm to fit the pipe nonlinear hysteresis, so as to obtain a pipe intrinsic nonlinear hysteresis response signal; The analysis steps of stripping the non-stationary background fluctuation include: The high-frequency pressure sequence is point-by-point corresponding to the pipe intrinsic nonlinear hysteresis response signal on the time axis; Based on the pipe intrinsic nonlinear hysteresis response signal after the subtraction operation, a preliminary residual is obtained, and the Hilbert-Huang transform is performed on the residual to decompose it into an intrinsic mode function; According to the Shannon entropy criterion, the intrinsic mode function is removed from the pipe slow-changing characteristic IMF component to reconstruct a pure fluid dynamics abnormal residual; The analysis steps of avoiding the modulation interference of the material nonlinearity on the fluid signal include: The operator density distribution characteristics of the P-I neuron and the multi-scale entropy characteristics of each IMF component are extracted respectively to obtain a first feature vector and a second feature vector, and the operator density distribution characteristics are determined according to the trained weight coefficient; Based on the first feature vector and the second feature vector, a high-dimensional joint feature vector is obtained, and a hysteresis-entropy conjugate feature mapping space is constructed; The analysis steps of avoiding the modulation interference of the material nonlinearity on the fluid signal also include: In the hysteresis-entropy conjugate feature mapping space, the high-dimensional joint feature vector is subjected to orthogonal subspace decomposition to construct a material memory main plane and a fluid disturbance orthogonal complement plane, and signal component vector decoupling is performed to obtain a material memory component vector and a fluid disturbance component vector; The two-norm of the fluid disturbance component vector is calculated by ratio with the sum of the two-norm of the material memory component vector and the two-norm of the fluid disturbance component vector to obtain a fluid dynamics confidence coefficient; The fluid dynamics confidence coefficient is used to perform soft threshold weighted reconstruction on the pure fluid dynamics abnormal residual to synthesize a faithful fluid dynamics abnormal residual; The analysis steps of determining the real leakage negative pressure wave and the sensor random drift also include: The Takens embedding theorem is used to reconstruct the faithful fluid dynamics abnormal residual into a high-dimensional phase space trajectory; Poincare section is used to intercept the continuous dynamic trajectory and convert it into a discrete two-dimensional point cloud distribution map; The discrete two-dimensional point cloud distribution map is input into a convolutional neural network for steady-state attractor pattern recognition and chaotic divergent trajectory analysis to determine the real leakage negative pressure wave and sensor random drift; wherein the steady-state attractor pattern recognition includes point-like or closed ring structure analysis of convergence, and the chaotic divergent trajectory analysis includes singular attractor or non-closed broken structure analysis of divergence; The step of taking the viscous-dominated memory decay cone as a geometric constraint specifically includes: For transient elastic impact vector clusters falling outside the viscous-dominated memory decay cone, orthogonal rejection or high-damping attenuation processing is performed on such vectors, the orthogonal rejection means mathematically projecting the vector to the direction perpendicular to the boundary of the cone and forcing it to be zero, or directly setting its amplitude to zero; the high-damping attenuation means multiplying the vector by a very small attenuation coefficient; For the steady-state viscous creep vector cluster falling in the view cone, its vector length is reserved, and a time weighting function based on the Boltzmann superposition principle is introduced to process it, and the calculation formula of the time weighting function is , wherein is a natural constant, represents the current calculation time, represents the historical time corresponding to the steady-state viscous creep vector, is a characteristic relaxation time constant of the material; all effective components after the above screening and weighting processing are subjected to vector superposition, the screening is to remove the vectors outside the view cone, and the weighting processing is to multiply the vectors inside the view cone by the time weighting function, that is, all the retained vectors after the time weighting are accumulated according to the parallelogram rule or the component summation method, and the generated composite vector is a pressure evolution string characteristic composite vector.

2. A method of intelligent leak detection for underground pipes as claimed in claim 1, wherein, The step of establishing the time-temperature equivalent equation calculation mobile factor specifically includes: selecting a standard reference temperature , using the Williams-Landel-Ferry equation or Arrhenius equation to calculate the mobile factor , the formula is , wherein: is the actual pipe wall temperature when collecting pressure data, is a logarithmic function, and are empirical constants related to the material of the pipe.

3. A method of intelligent leak detection for underground pipes as claimed in claim 1, wherein, The step of inverting the material relaxation time spectrum based on the temperature-eliminated equivalent pressure sequence specifically includes: Using the mobile factor, the high-frequency pressure sequence collected on the physical time axis is mapped onto the reduced time axis . The calculation formula is . The calculation formula is , where represents the current physical observation time, is the integral variable, represents the instantaneous temperature at the historical time ; the equivalent pressure sequence is regarded as a known relaxation modulus evolution sequence, a series of discrete relaxation time points are set, and the relaxation modulus function is fitted by nonlinear least squares , where represents the relaxation modulus changing with time, represents the equilibrium modulus, represents the total number of parallel Maxwell units in the generalized Maxwell model; is the summation index, representing the th Maxwell unit, represents the modulus strength of the th relaxation unit, represents the characteristic relaxation time point of the th relaxation unit, represents the exponential function operation with the natural constant as the base; the modulus strength corresponding to each relaxation time point is arranged according to the relaxation time distribution, and the material relaxation spectrum density matrix in the form of a diagonal matrix is constructed .

4. The intelligent leak detection method for underground pipes according to claim 1, wherein, The step of performing dynamic geometric chord intersection, connecting adjacent wave amplitude point sets to form pressure evolution chord vectors, and marking the evolution chord vectors in geometric space as transient elastic impact vector clusters and steady-state viscous creep vector clusters according to the slope of the pressure evolution chord vector and the inherent elastic-viscous critical threshold of the pipe specifically includes: Connect two adjacent fluctuation amplitude points sequentially according to the time series. and ,in For the first The fluctuation amplitude relative to the baseline at each moment is used to construct a directed line segment, i.e., a pressure evolution chord vector. Calculate the slope of each chord vector. , Indicates the first The timestamps of each sampling point; the inherent elastic-viscosity critical threshold of the pipe material. These are physical parameters obtained through dynamic thermomechanical analysis of standard pipe samples, and the calculated absolute value of the slope is... With threshold Comparison: If This vector is labeled as a transient elastic impact vector cluster; if This vector is labeled as a steady-state viscous creep vector cluster.

5. A method of intelligent leak detection for underground pipes as claimed in claim 4 wherein, The step of weighting the influence weight of different historical time on the current state through the multi-head attention mechanism to obtain a weighted fusion feature vector specifically includes: The full history stress-strain memory field tensor Refers to a two-dimensional matrix constructed, containing time dimension And relaxation spectrum dimension, extract features through multi-head attention mechanism, set the number of attention heads as h, for the first Attention head, map the input tensor to query vector , key vector And value vector Through linear transformation, calculate using the scaled dot-product attention formula , which Represents the query feature of the current state, The index key value of the historical moment, Indicates the transpose of the index key value of the historical moment, Carries specific residual stress numerical information, The dimension of the key vector is Indicates the attention score matrix, Softmax is the normalization exponential function, the score matrix is multiplied by the value vector To realize the weighted sum of the full history residual stress information, concatenate and fuse the outputs of all The heads of the linear transformation output the weighted fusion feature vector.

6. The intelligent leak detection method for underground pipes according to claim 1, wherein, The step of extracting the operator density distribution features of the P-I neuron specifically includes: Based on the trained P-I neural network, a parameter set representing the memory characteristics of the pipe material is extracted, i.e. for the number of parallel P-I neurons , a set of weight coefficients is obtained after training convergence , wherein , the set of weight coefficients is arranged in the order of neuron index to form a -dimensional vector, which is defined as the operator density distribution characteristics of the P-I neuron.

7. The intelligent leak detection method for underground pipes according to claim 1, wherein, The step of intercepting Poincare section in phase space, converting continuous dynamic trajectory into discrete two-dimensional point cloud distribution map, inputting the discrete two-dimensional point cloud distribution map into convolutional neural network, and performing steady-state attractor pattern recognition and chaotic divergent trajectory analysis to determine the real leakage negative pressure wave and sensor random drift specifically includes: In the constructed In the phase space, a plane formed by the two principal component directions with the largest variability in the phase space trajectory is selected as the cross-section projection reference, or a plane passing through the geometric center of the phase space trajectory and perpendicular to the average flow direction of the trajectory is selected as the Poincare cross-section; all one-way crossing points of the high-dimensional phase space trajectory and the Poincare cross-section are recorded to obtain a discrete intersection set; the high-dimensional intersection points are mapped to a two-dimensional plane coordinate system to generate a two-dimensional scatter plot, and the two-dimensional scatter plot is subjected to binary processing and normalized mapping to generate a black and white binary image, which is a discrete two-dimensional point cloud distribution diagram; A lightweight convolutional neural network model is constructed, which includes an input layer for receiving a point cloud pixel matrix, a plurality of convolutional layers for extracting point cloud spatial distribution features, a pooling layer for dimension reduction and feature invariance, and a fully connected layer; a training data set is established and a topological morphology label is defined, Poincare section point cloud images are generated by collecting historical data, and the point cloud is divided into normal or interference class labels and leakage class labels according to the geometric topological features of the point cloud; The CNN model is trained under supervised supervision using the labeled dataset. Forward propagation is used to calculate the prediction results, and the cross-entropy loss function is employed to quantify the error between the predicted values ​​and the true mathematical topological labels. Backpropagation is used to calculate the gradient, and the network weights are iteratively updated using an optimizer. A discrete two-dimensional point cloud distribution map is input into the trained CNN model. The model's output layer uses the Softmax function to calculate the confidence probability of the input image belonging to each category. If the model outputs the probability of belonging to the leak class... If the alarm value is greater than or equal to the preset alarm threshold, a real leak is determined to have occurred within the fluid system; otherwise, if... If the noise level is below the threshold, it is determined to be background noise or periodic interference, and the system has no leakage.

8. An intelligent leak detection apparatus for underground pipes, characterized in that, A method for intelligent leakage detection of underground pipelines according to any one of claims 1-7, comprising: A memory conversion module is used to calculate a shift factor based on the viscoelastic rheological response law of the pipe material, map the high-frequency pressure sequence on the physical time axis to a reduced time axis using the shift factor, and obtain a temperature-eliminated equivalent pressure sequence; and convert the temperature-eliminated equivalent pressure sequence into a full-history stress-strain memory field tensor; A hysteresis response module is used to construct a neural hysteresis operator network, input the full-history stress-strain memory field tensor, analyze the non-local memory characteristics and hysteresis loop, and obtain the intrinsic nonlinear hysteresis response signal of the pipe material; An abnormal residual module is configured to perform point-by-point difference processing on the high-frequency pressure sequence and the pipe material intrinsic nonlinear hysteresis response signal, strip non-stationary background fluctuations, avoid modulation interference of material nonlinearity on the fluid signal, and extract a faithful fluid dynamics abnormal residual; A leakage judgment module is configured to reconstruct the faithful fluid dynamics abnormal residual into a high-dimensional phase space trajectory, construct a spatiotemporal evolution Poincare section, input into a convolutional neural network, and judge a real leakage negative pressure wave and sensor random drift through an attractor shape difference on the spatiotemporal evolution Poincare section.

Images