Urban pipeline network health status evaluation method based on artificial intelligence

Abstract

The present invention relates to the field of measuring electrical variables, and specifically to an artificial intelligence-based method for evaluating the health status of urban pipe networks. The method comprises obtaining structural parameters of the urban pipe network and measuring electrical variable data; synchronously acquiring working electrode voltage and reference electrode voltage, and outputting a compensated voltage value; calculating an adjusted signal based on a gain coefficient; synchronously acquiring positive and negative phase signals, performing a differential operation, and obtaining an anti-interference output vector; constructing a differential entropy feature matrix; constructing and labeling a training data set; obtaining a trained health status evaluation model; obtaining data to be measured, inputting the data to be measured into the trained health status evaluation model, and the trained health status evaluation model outputting the probability of each health status label, with the category with the highest probability being the evaluation result. Existing methods have a low accuracy rate in evaluating the health status of urban pipe networks. The present method has a high accuracy rate in evaluating the health status of urban pipe networks.

Description

Technical Field

[0001] The present invention relates to the field of measuring electrical variables, and in particular to an artificial intelligence-based method for evaluating the health status of urban pipe networks. Background Art

[0002] Urban pipeline networks are an integral part of modern urban infrastructure, carrying the crucial task of transporting resources such as water, electricity, and gas. Urban pipeline networks are often located underground and are complex, making their health difficult to monitor. The health of urban pipeline networks is directly linked to the smooth functioning of cities and the quality of life of residents. Therefore, monitoring their health is crucial, using methods such as inspections or real-time measurement and analysis using sensors combined with artificial intelligence (AI) to prevent resource waste and serious safety incidents.

[0003] In urban pipe networks, pipes used for transporting liquids, such as water supply, drainage, and heating, are more susceptible to corrosion and leakage. Existing methods for evaluating the health status of urban pipe networks have the following problems:

[0004] (1) The constant voltage measurement method is often used to measure the conductivity of urban pipe networks. During the measurement, a constant voltage is directly applied and the current is measured. Therefore, the dynamic changes of contact impedance are ignored and the true conductivity cannot be accurately obtained. As a result, the measurement results are easily affected by factors such as environmental humidity or corrosive substances, resulting in baseline drift, which affects the accurate assessment of the health status of urban pipe networks.

[0005] (2) The conductivity of different sections of the urban pipeline network varies significantly. The existing method uses a segmented gain strategy to switch the gain gear according to the preset threshold. This cannot adapt to the rapid dynamic changes of conductivity, resulting in signal distortion, easy signal saturation in low-conductivity areas, and insufficient resolution in high-conductivity areas. This affects the accuracy and reliability of signal measurement, thereby affecting the accurate assessment of the health status of the urban pipeline network.

[0006] (3) There are power frequency interference and harmonic pollution on the urban pipeline network. The existing methods use conventional processing methods such as band-stop filters to suppress power frequency noise. However, while suppressing the noise, the high-frequency components of the fault characteristics will be attenuated, resulting in the masking of weak fault characteristics such as pipeline corrosion or leakage, reducing the detection sensitivity and failing to effectively retain the high-frequency components of the fault characteristics, thus affecting the accurate assessment of the health status of the urban pipeline network.

[0007] (4) When extracting fault features from electrical variable data, we rely on time-domain statistics such as mean and variance. These statistics are not sensitive enough to non-Gaussian distribution fault features, making it difficult to effectively capture the energy distribution characteristics and complexity information of the signal, resulting in insufficient fault feature extraction, thereby eliminating the impact of power frequency interference on the accurate assessment of the health status of urban pipeline networks. At the same time, the health status evaluation model of existing technologies may not fully preserve the amplitude-phase coupling relationship between features, and lack the ability to respond nonlinearly to urban pipeline network fault features, affecting the accuracy of health status evaluation.

[0008] Therefore, the existing urban pipeline network health status evaluation method has the problem of low assessment accuracy of the urban pipeline network health status. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide an urban pipeline network health status evaluation method based on artificial intelligence with high accuracy in evaluating the health status of urban pipeline networks.

[0010] To solve the above technical problems, the present invention provides an artificial intelligence-based urban pipe network health status evaluation method comprising:

[0011] S1. Obtain urban pipe network structural parameters and measure electrical variable data;

[0012] S2. Synchronously acquire the working electrode voltage and the reference electrode voltage based on the four-electrode differential measurement method; utilize a dynamic compensation mechanism to output the compensated voltage value;

[0013] S3. Calculate the signal fluctuation intensity and dynamically adjust the gain coefficient, and calculate the adjusted signal based on the gain coefficient;

[0014] S4. synchronously collect the positive-phase and negative-phase signals, perform differential operations, apply a frequency-domain Gaussian mask to suppress noise, and obtain an anti-interference output vector;

[0015] S5. Construct an analytical signal through Hilbert transform and multi-scale wavelet decomposition, calculate the energy probability distribution and differential entropy characteristics, and build a differential entropy feature matrix;

[0016] S6. Build and label the training dataset;

[0017] S7. constructing a health status evaluation model using an end-to-end complex neural network architecture, training the health status evaluation model, and obtaining a trained health status evaluation model;

[0018] S8. After executing S1 to S5 in the urban pipe network test area, obtain the test data, input the test data into the trained health status evaluation model, and the trained health status evaluation model outputs the probability of each health status label, and takes the category with the highest probability as the evaluation result.

[0019] As a further improvement of the present invention: the step of S1 includes:

[0020] S101. Obtaining urban pipe network structural parameters:

[0021] S102. Synchronously measure electrical variable data;

[0022] Install four-electrode sensing units at key nodes of the urban pipe network to achieve high-precision acquisition of electrical variable data;

[0023] Electrical variable data include: conductivity raw signal, current and voltage related variables; current and voltage related variables include excitation current, working electrode voltage vector, reference electrode voltage vector;

[0024] Key nodes in urban pipeline networks include flanges, elbows, and valves.

[0025] As a further improvement of the present invention: the step of S2 includes:

[0026] S201. Based on the four-electrode differential measurement principle, a constant current is injected while the working electrode voltage and the reference electrode voltage are simultaneously acquired. A dynamic compensation mechanism is used to eliminate baseline drift caused by the electrode-tube wall interface impedance using a calibration resistor and reference loop impedance, and the compensated voltage value is output.

[0027] S202. Under no-load conditions without connection to the urban pipeline network, inject a calibration current and measure the no-load reference voltage. The calibration resistance is calculated by the ratio of the calibration current to the no-load reference voltage and used as a constant in the dynamic compensation formula.

[0028] As a further improvement of the present invention: the step of S3 includes:

[0029] S301. Taking the current sampling point as the center, select the historical compensation voltage data points contained in the sliding window, and calculate the average value of the selected historical compensation voltage data points as the local voltage reference value;

[0030] S302. Calculate the standard deviation of the compensation voltage data within the sliding window, combine the range scaling factor and the anti-zero constant, and determine the adaptive gain coefficient;

[0031] S303. Multiply the gain coefficient of the current sampling point by the compensation voltage value, and output the adjusted signal.

[0032] As a further improvement of the present invention: the step of S4 includes:

[0033] S401. In a single excitation cycle, synchronously collect the positive phase voltage signal and the negative phase voltage signal to form a signal pair for subsequent differential processing;

[0034] S402. Perform differential operation on the positive and negative phase signals within the same excitation cycle and take the absolute value of the result;

[0035] S403. Averaging the differential absolute value signals of multiple consecutive excitation cycles to reduce the influence of random noise, converting the averaged signal into the frequency domain through fast Fourier transform to obtain a frequency domain representation;

[0036] S404. Multiply the frequency domain Gaussian mask and the average frequency domain signal element by element to obtain a filtered frequency domain signal;

[0037] S405. Perform inverse Fourier transform on the filtered frequency domain signal, convert it back to the time domain, and output an anti-interference output vector.

[0038] As a further improvement of the present invention: the step of S5 includes:

[0039] S501. Performing a Hilbert transform on the input anti-interference output signal to generate a complex analytical signal containing amplitude and phase information;

[0040] S502. Perform multi-scale decomposition of the complex analytic signal using the Morlet wavelet basis function, and extract wavelet components of different frequency bands at preset high, medium, and low frequency scales;

[0041] S503. For all wavelet components at each scale, calculate the ratio of the energy of each component to the total energy of the scale to form an energy probability distribution:

[0042] S504. Calculate differential entropy features based on the energy probability distribution of each scale;

[0043] S505. Integrate the differential entropy features of the three preset scales of high frequency, medium frequency and low frequency to form a differential entropy feature matrix.

[0044] As a further improvement of the present invention: the step of S7 includes:

[0045] S701. Define the health status evaluation model structure;

[0046] The health status evaluation model adopts an end-to-end complex neural network architecture, with the input being the differential entropy feature matrix , the output is the probability distribution of the health status of the urban pipeline network;

[0047] S702. Constructing an adaptive feature gating layer based on the adaptive feature gating mechanism;

[0048] S703. Constructing a dual-domain activation function layer based on the dual-domain activation function;

[0049] S704. Construct a gradient-stabilized batch normalization layer based on gradient-stabilized batch normalization;

[0050] S705. Output the classification result through the fully connected classification layer;

[0051] S706. Calculate blur boundary focus loss;

[0052] S707. Repeat iterations S702 to S705 until a preset stop iteration condition is met.

[0053] Preferably, the step of S703 includes:

[0054] S7031. Perform a modulus operation on the enhanced feature matrix to extract the amplitude information of each element and generate a feature modulus matrix;

[0055] S7032. Perform phase angle calculation on the enhanced feature matrix to extract the phase direction information of each element to obtain a feature phase angle matrix;

[0056] S7033. Input the eigenvalue matrix into the trainable weight matrix, perform nonlinear transformation through the Sigmoid activation function, and multiply the transformation result element-by-element with the original eigenvalue matrix;

[0057] S7034. Input the phase angle matrix into the trainable weight matrix, perform nonlinear transformation using the hyperbolic tangent function, and multiply the transformation result element-by-element with the original phase angle matrix;

[0058] S7035. Use the transformed eigenvalue matrix as the real part and the transformed phase matrix as the imaginary part to generate a complex eigenmatrix.

[0059] Preferably, the step of S704 includes:

[0060] S7041. Decompose the input complex feature matrix into independent real and imaginary matrices;

[0061] S7042. Based on the current training batch data, calculate the mean and variance of all elements of the real matrix;

[0062] S7043. Use the real batch statistics to normalize the real matrix, restore the feature expression ability through the trainable scaling parameters and translation parameters, and output the normalized real matrix;

[0063] S7044. Based on the current training batch data, calculate the mean and variance of all elements of the imaginary matrix;

[0064] S7045. Normalize the imaginary part matrix using the imaginary part batch statistic, adjust the imaginary part feature distribution through independent trainable scaling parameters and translation parameters, and output the normalized imaginary part matrix;

[0065] S7046. Recombine the normalized real matrix and imaginary matrix into a complex characteristic matrix.

[0066] Preferably, the step of S706 includes:

[0067] S7061. Calculate the fuzzy membership of the sample to each category based on the Euclidean distance between the true sample label and the pre-clustered cluster center and the variance of the intra-class distribution;

[0068] S7062. Calculate the inverse frequency square root weight based on the frequency of occurrence of each category in the training set;

[0069] S7063. Construct modulation factors based on model prediction probabilities;

[0070] S7064. Combining fuzzy membership, class weight factor and focus modulation factor to weight the cross entropy loss of each class;

[0071] S7065. Sum the weighted loss components of all categories to generate the fuzzy boundary focus loss.

[0072] The beneficial effects of the present invention are as follows: The urban pipeline network health status evaluation method based on artificial intelligence provided by the present invention has a high accuracy rate in evaluating the health status of urban pipeline networks.

[0073] First, based on the four-electrode differential measurement method, S2 synchronously collects the working electrode voltage and reference electrode voltage data by injecting excitation current, and uses a dynamic compensation mechanism to eliminate the influence of the electrode-pipe wall interface impedance. It can accurately measure conductivity, avoid baseline drift problems, and ensure that the conductivity measurement results accurately reflect the true state of the pipeline network, thereby providing a reliable basis for the accurate assessment of the health status of the pipeline network; S3 monitors the fluctuation characteristics of the compensation voltage in real time, calculates the signal fluctuation intensity, and dynamically adjusts the gain coefficient to stabilize the output signal within the target dynamic range. It can adapt to the rapid dynamic changes of conductivity in different sections of the urban pipeline network, and effectively solves the problem that fixed-gain amplifiers are prone to signal saturation in low conductivity areas and high conductivity areas. The S4 algorithm addresses the problem of insufficient resolution and improves the accuracy and reliability of signal measurement. It uses a hybrid mechanism combining time-domain correlated double sampling circuits and frequency-domain selective filtering to synchronously acquire positive and negative phase signals for differential calculation to eliminate power frequency interference. It also applies a frequency-domain Gaussian mask to suppress noise while retaining the high-frequency components of fault characteristics. This overcomes the drawback of conventional band-stop filters, which attenuate high-frequency components of fault characteristics when suppressing power frequency noise, masking weak fault characteristics such as pipeline corrosion or leakage and reducing detection sensitivity. The S5 algorithm constructs an analytical signal using Hilbert transform and multiscale wavelet decomposition to obtain amplitude and phase information, extract components in different frequency bands, calculate energy probability distributions and differential entropy features, and construct a differential entropy feature matrix. Furthermore, an end-to-end complex neural network architecture is used to construct and train a health status assessment model. The trained model retains the amplitude-phase coupling relationship between features and optimizes the nonlinear response capability for pipeline network fault characteristics, improving the adequacy of fault feature extraction and the accuracy of health status assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a principle block diagram of the present invention;

[0075] Figure 2 This is the experimental diagram of measurement error of different impedance compensation methods under humidity changes;

[0076] Figure 3 This is a comparison experiment diagram of measurement error distribution;

[0077] Figure 4 This is the signal quality experimental diagram of the dynamic gain adjustment strategy under different conductivities;

[0078] Figure 5 This is the experimental diagram of the adaptive gain adjustment strategy;

[0079] Figure 6 This is a performance comparison experiment diagram for evaluating the health status of different models;

[0080] Figure 7 This is the experimental diagram of the model confusion matrix. DETAILED DESCRIPTION

[0081] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0082] like Figure 1 As shown, the urban pipe network health status evaluation method based on artificial intelligence provided by the present invention includes:

[0083] S1. Obtain urban pipe network structural parameters and measure electrical variable data;

[0084] S101. Obtaining urban pipe network structural parameters:

[0085] Consult urban pipe network design drawings or as-built documents and other documents containing pipeline parameter information to obtain the geometric dimensions and physical properties of the urban pipe network;

[0086] The geometric dimensions and physical properties of the urban pipe network include pipe inner diameter, pipe wall thickness, and pipe material code (metal or non-metal);

[0087] For pipes located above ground in urban pipe networks, a GPS locator can be used to record the three-dimensional spatial position, an ultrasonic caliper can be used to measure the inner diameter of the pipe, and an ultrasonic probe can be used to measure the wall thickness.

[0088] When measuring the inner diameter of a pipe using an ultrasonic caliper, measure once every 1.0m along the pipeline and take the average of three measurements as the measured value. When measuring the pipe wall thickness using an ultrasonic probe, measure three points at the flange connection and the middle of the pipe body and calculate the average value as the measured value.

[0089] S102. Synchronously measure electrical variable data;

[0090] Install four-electrode sensing units at key nodes of the urban pipe network to achieve high-precision acquisition of the original conductivity signal in the electrical variable data;

[0091] Electrical variable data include: conductivity raw signal, current and voltage related variables; current and voltage related variables include excitation current, working electrode voltage vector, reference electrode voltage vector;

[0092] Key nodes of urban pipe networks include flanges, elbows, valves, etc.

[0093] The arrangement rules of the electrodes of the four-electrode sensing unit are:

[0094] Working electrodes (WE1, WE2): stainless steel needle electrodes, embedded 2 mm deep into the inner surface of the tube wall;

[0095] Reference electrode (RE1, RE2): Ag / AgCl reference electrode, 10 cm away from the working electrode.

[0096] S2. Synchronously acquire the working electrode voltage and the reference electrode voltage based on the four-electrode differential measurement method; utilize a dynamic compensation mechanism to output the compensated voltage value;

[0097] S201. Calculate contact impedance compensation voltage

[0098] Based on the four-electrode differential measurement principle, a constant current is injected and the working electrode voltage and reference electrode voltage are synchronously collected. Through a dynamic compensation mechanism, the calibration resistor and reference loop impedance are used to eliminate the baseline drift caused by the electrode-pipe wall interface impedance. The compensated voltage value is output to eliminate the influence of the electrode-pipe wall interface impedance. The output compensation voltage value is a real vector, ensuring that the conductivity measurement results accurately reflect the actual status of the urban pipe network.

[0099] Injecting a constant current and synchronously collecting the working electrode voltage and the reference electrode voltage can be specifically done by injecting a constant current excitation into WE1–WE2. , representing a square wave with a frequency of ;Synchronously collect the measured voltage vector of the working electrode and the reference electrode voltage vector , the sampling rate is 1kHz, and the total number of sampling points is N;

[0100] Compensation voltage value The calculation formula is:

[0101] ,

[0102] Where, is the measured voltage vector of the working electrode, representing the voltage measurement value at the contact point between the electrode and the urban pipe network; is the reference electrode voltage vector, representing the voltage measurement value of the reference point; is the calibration resistor, which is a real scalar; is the reference loop impedance, which is a real scalar and represents the equivalent impedance of the reference electrode path.

[0103] Reference loop impedance The test current is measured by applying a known test current under no-load conditions. And measure the reference voltage , reference loop impedance The calculation formula is:

[0104] ,

[0105] In the formula, if the reference voltage , reference loop impedance ,but ,Right now .

[0106] S202. Calculate the calibration resistance;

[0107] In the no-load condition without urban pipe network connection, inject the calibration current And measure the no-load reference voltage , by calibrating the ratio of the current to the no-load reference voltage, the calibration resistance is calculated and used as a constant in the dynamic compensation formula to ensure that the compensation voltage calculation (i.e., the formula ) accuracy;

[0108] Calibration resistor Calculated by initial calibration, calibration resistance The calculation formula is:

[0109] ,

[0110] Where, is the no-load reference voltage, which is a real scalar and represents the initial value of the reference electrode voltage when there is no load; is the calibration current, which is a real scalar and represents the initial calibration value of the injected current; if the no-load reference voltage , calibration current ,but ,Right now .

[0111] When measuring the conductivity of urban pipe networks, electrode contact impedance is easily affected by ambient humidity or corrosive substances, leading to baseline drift and compromising measurement accuracy. Conventional approaches employ constant voltage measurement, directly applying a constant voltage and measuring the current. However, this ignores the dynamic changes in contact impedance and fails to accurately capture true conductivity, thus impacting assessments of the health of urban pipe networks. The S2, based on a four-electrode differential measurement method, injects an excitation current into the urban pipe network, synchronously collects working and reference electrode voltage data, and utilizes a dynamic compensation mechanism to eliminate the influence of the electrode-pipe wall interface impedance, ensuring that conductivity measurements accurately reflect the true state of the urban pipe network and avoiding baseline drift. To evaluate the measurement stability of different impedance compensation methods under varying ambient humidity conditions, the conventional constant voltage measurement method, the fixed compensation method, and the dynamic compensation mechanism of the present invention were compared to experimentally analyze the error variation in conductivity measurement within a humidity range of 30%-95%. The conventional method showed a significant increase in error with increasing humidity. In particular, in humidity environments above 70%, changes in electrode contact impedance led to a 26% error peak in the constant voltage measurement method. Although the fixed compensation method showed some improvement, it still had an error fluctuation of 15%. The present invention, based on the dynamic compensation mechanism of four-electrode differential measurement, stabilized the measurement error within 3% through real-time calibration of the resistance and reference loop impedance adjustment. Figure 2 The curve shows that the error bar of the present invention is nearly horizontal. Figure 3The box plot further confirms that the error distribution range is narrow (concentrated in 1%-3%), proving that the present invention effectively solves the problem of impedance drift at the electrode-pipe wall interface and provides a reliable data basis for urban pipe network health assessment.

[0112] S3. Calculate the signal fluctuation intensity and dynamically adjust the gain coefficient, and calculate the adjusted signal based on the gain coefficient;

[0113] S301. Calculate the sliding window mean;

[0114] Taking the current sampling point as the center, select the historical compensation voltage data points contained in the sliding window, and calculate the average value of the selected historical compensation voltage data points as the local voltage reference value. The local voltage reference value is a real number used to represent the average level of the local voltage of the current signal;

[0115] Sliding window mean at the kth sampling point The calculation formula is:

[0116] ,

[0117] Where W is the sliding window width, which is a positive integer and defines the number of historical data points included in the mean calculation. It can be set to W=50; i is a positive integer; k represents the time series index of the current sampling point, which is a positive integer and is used to identify the position in the signal sequence. It can be set to Corresponding to the 1st, 2nd, 3rd, ... sampling points respectively; summation interval To k means taking the current point and the previous historical points; is the compensation voltage value of the i-th sampling point, which is a real number;

[0118] S302. Calculate the gain coefficient;

[0119] The standard deviation of the compensated voltage data within the sliding window is calculated to characterize the signal fluctuation intensity. The adaptive gain coefficient is determined by combining the range scaling factor and the zero-prevention constant. The adaptive gain coefficient is a real number and is used to dynamically adjust the signal amplitude to ensure that the output is stable within the target dynamic range.

[0120] Gain coefficient of the kth sampling point The calculation formula is:

[0121] ,

[0122] Where, is the range scaling factor, which is a real constant and controls the target amplitude range of the output signal. =2.5; The term represents the variance of the compensation voltage within the sliding window, which characterizes the intensity of signal fluctuation; To prevent the division of zero constants, which are extremely small positive real numbers and avoid numerical anomalies with denominators of zero, we can set .

[0123] S303. Apply gain adjustment;

[0124] Multiply the gain coefficient of the current sampling point by the compensation voltage value and output the adjusted signal. The adjusted signal is a real number, ensuring that the signal amplitude of different conductivity segments always remains within the preset target voltage range;

[0125] Voltage value of the kth sampling point excitation cycle The calculation formula is:

[0126] ,

[0127] Where, is the compensation voltage value of the kth sampling point.

[0128] The conductivity of different sections of the urban pipe network varies significantly, which causes fixed-gain amplifiers to easily saturate signals in low-conductivity areas and have insufficient resolution in high-conductivity areas. Conventional processing methods use segmented gain strategies to switch gain levels based on preset thresholds, but the response is delayed and cannot adapt to the rapid dynamic changes in conductivity, resulting in signal distortion. S3 monitors the fluctuation characteristics of the compensation voltage in real time, calculates the signal fluctuation intensity, and dynamically adjusts the gain coefficient to ensure that the output signal is stable within the target dynamic range, avoiding signal saturation or insufficient resolution, and adapting to rapid changes in different conductivity sections. To verify the adaptability of different gain strategies to dynamic changes in conductivity, a wide range of conductivity changes from 0.1 to 100 Siemens / meter was tested. Figure 4 The data shows that when comparing the signal quality of fixed gain, segmented gain, and S3 adaptive gain adjustment, the fixed gain has a signal-to-noise ratio of less than 10 dB in the low conductivity area (<1 Siemens / m) due to weak signals. In the high conductivity area (>50 Siemens / m), the signal-to-noise ratio deteriorates due to signal saturation. The segmented gain has a clear signal-to-noise ratio gap at the thresholds of 20 Siemens / m and 40 Siemens / m. S3 calculates the signal fluctuation intensity in real time through a sliding window and dynamically adjusts the gain coefficient. Figure 5 The display gain changes smoothly with conductivity, keeping the signal-to-noise ratio above 22 decibels over the entire range. Especially in the extreme conductivity area, S3 improves signal quality by 40% compared with conventional methods, solving the signal distortion problem caused by response lag in traditional methods.

[0129] S4. A hybrid mechanism of time-domain correlated double sampling circuitry and frequency-domain selective filtering is used to synchronously acquire positive and negative phase signals for differential calculation to eliminate power frequency interference. A frequency-domain Gaussian mask is then applied to suppress noise while retaining the high-frequency components characteristic of the fault.

[0130] S401. Synchronous acquisition of positive and negative phase signals

[0131] In a single excitation cycle, the positive phase voltage signal and the negative phase voltage signal are synchronously collected to form a signal pair for subsequent differential processing, which is expressed as:

[0132] ,

[0133] Where, Represents the positive phase voltage signal vector of the excitation cycle of the mth sampling point, which belongs to the real vector space; The inverted voltage signal vector representing the excitation cycle of the mth sampling point belongs to the real vector space; m is a positive integer; M is the total number of cycles, representing the accumulated number of excitation cycles, and M can be set to 8.

[0134] S402. Calculate the absolute value of the period difference;

[0135] Perform differential operation on the positive and negative phase signals within the same excitation cycle and take the absolute value of the result to eliminate the common mode component of the power frequency interference while retaining the signal amplitude characteristics, which is expressed as:

[0136] ,

[0137] Where, Represents the absolute value vector of the difference in the mth period, belongs to the real vector space, and its dimension is the number of sampling points in a single period; is the absolute value operation, The absolute value operation of the term is used to eliminate the influence of signal polarity;

[0138] S403. Multi-period averaging and frequency domain conversion;

[0139] The differential absolute value signals of multiple consecutive excitation cycles are averaged to reduce the influence of random noise, and the averaged signal is converted to the frequency domain through fast Fourier transform to obtain the frequency domain representation, which belongs to the complex vector space;

[0140] Frequency domain representation of the average difference signal The calculation formula is:

[0141] ,

[0142] Where, represents fast Fourier transform;

[0143] S404. Apply frequency domain Gaussian mask;

[0144] The frequency domain Gaussian mask is multiplied element-by-element by the average frequency domain signal to eliminate the power frequency interference and suppress the power frequency interference while retaining the high-frequency fault characteristic components outside the frequency band, which can be expressed as:

[0145] ,

[0146] Where, Represents the frequency domain signal after filtering, which belongs to the complex vector space; It is an element-by-element multiplication operation; is the frequency domain Gaussian mask vector;

[0147] The calculation formula is:

[0148] ,

[0149] Where, e is a natural constant; is the nth frequency point corresponding to the fast Fourier transform; N is the total number of sampling points; n is a positive integer; if the frequency domain Gaussian mask vector The gain is minimum at 50Hz, that is , the gain is close to 1 outside the 40-60Hz range, thereby eliminating power frequency interference and effectively suppressing the 40-60Hz power frequency interference band, while retaining high-frequency fault characteristics outside this frequency band, such as harmonic components of pipeline corrosion and leakage;

[0150] S405. Inverse transform to generate anti-interference output;

[0151] Perform inverse Fourier transform on the filtered frequency domain signal, convert it back to the time domain, and output the anti-interference output vector, which is expressed as:

[0152] ,

[0153] Where, represents the anti-interference output vector, which belongs to the real vector space; represents the inverse fast Fourier transform, which converts the frequency domain signal back to the time domain;

[0154] Urban pipeline networks are subject to power frequency interference and harmonic contamination. Conventional treatment methods, such as band-stop filters, use filtering to suppress power frequency noise. However, this noise suppression attenuates the high-frequency components of fault signatures, masking subtle fault characteristics such as pipeline corrosion or leakage, and reducing detection sensitivity. The S4 employs a hybrid mechanism combining time-domain correlated double sampling circuitry and frequency-domain selective filtering. By synchronously sampling positive and negative phase signals within the excitation cycle, performing differential operations to eliminate power frequency interference, and applying frequency-domain filtering to suppress noise, the S4 eliminates power frequency interference, effectively retains the high-frequency components of fault signatures, and simultaneously suppresses power frequency interference, improving anti-interference capabilities.

[0155] S5. Analytical signals are constructed using Hilbert transform and multi-scale wavelet decomposition to obtain amplitude and phase information. Components in different frequency bands are extracted and their energy probability distribution and differential entropy features are calculated to construct a differential entropy feature matrix. Furthermore, an end-to-end complex neural network architecture is used to construct and train a health status assessment model. The trained health status assessment model retains the amplitude-phase coupling relationship between features and optimizes the nonlinear response capability based on pipeline network fault characteristics.

[0156] S501. Construct analytical signal

[0157] Perform Hilbert transform on the input anti-interference output signal to generate a complex analytical signal containing amplitude and phase information, which belongs to the complex vector space;

[0158] The analytical signal at time t The calculation formula is:

[0159] ,

[0160] Where, is the sampling value of the anti-interference output vector at the tth moment, which is a real number; is an imaginary unit, that is ; represents the Hilbert transform operator, defined as , increasing the quadrature component by 90° phase shift; t is a positive integer.

[0161] S502. Multiscale wavelet decomposition;

[0162] The Morlet wavelet basis function is used to perform multi-scale decomposition on the complex analytical signal. Wavelet components of different frequency bands are extracted at the preset high, medium and low frequency scales. The wavelet components are complex numbers.

[0163] The sth scale wavelet components The calculation formula is:

[0164] ,

[0165] Where, represents the inner product operation; is the translation-scaled Morlet wavelet basis function;

[0166] The calculation formula is:

[0167] ,

[0168] Where s is the wavelet decomposition scale, and its value set is , corresponding to high-frequency, medium-frequency, and low-frequency components respectively; is the kernel function of the Morlet wavelet basis function, and its input is defined as is the normalized time variable , the calculation method is expressed as ; is the center frequency, which can be set =6; is the component index, and , is the total number of components at a specified scale, which can be set =3;

[0169] S503. Calculate component energy probability;

[0170] For all wavelet components at each scale, the proportion of the energy of each component to the total energy of the scale is calculated to form an energy probability distribution, which reflects the distribution characteristics of signal energy in different frequency bands and is expressed as:

[0171] ,

[0172] Where, Indicates the scale s The energy probability of each component is a real number; It is a modulo operation; express The model; Represents the total energy of scale s; i is a positive integer; is the total number of components; represents the i-th wavelet component of the s-th scale;

[0173] The calculation formula is:

[0174] ,

[0175] Where, represents the real part; Indicates the sth scale wavelet components; Indicates taking the imaginary part;

[0176] S504. Calculate differential entropy characteristics;

[0177] The differential entropy feature is calculated based on the energy probability distribution at each scale. The differential entropy feature is a real number that represents the uncertainty and complexity of the signal in the corresponding frequency band and is used to quantify the fault sensitivity characteristics.

[0178] Differential entropy characteristics of scale s The calculation formula is:

[0179] ,

[0180] Where, For natural logarithm operation, the default base is e.

[0181] S505. Construct a feature matrix;

[0182] The differential entropy features of three preset scales, high frequency, medium frequency and low frequency, are integrated to form a differential entropy feature matrix, which represents the key features of the health status of the urban pipeline network and is expressed as:

[0183] ,

[0184] Where, represents the differential entropy characteristic matrix, which belongs to the 3-dimensional real vector space; represents the differential entropy feature with a scale of 1; represents the differential entropy feature with a scale of 3; Represents the differential entropy feature with a scale of 5.

[0185] Non-Gaussian distribution fault signatures in electrical variable data are insufficiently sensitive to time-domain statistics such as mean and variance. Conventional processing methods rely on these statistics to characterize the signal, making it difficult to effectively capture the signal's energy distribution characteristics and complexity, resulting in inadequate fault feature extraction. S5 uses the Hilbert transform and multi-scale wavelet decomposition to first construct an analytical signal to obtain amplitude and phase information. It then performs multi-scale decomposition to extract components in different frequency bands and calculates energy probability distribution and differential entropy features. This eliminates power frequency interference and forms a differential entropy feature matrix that characterizes fault sensitivity, quantifying signal uncertainty and complexity.

[0186] S6. Build and label the training dataset;

[0187] Through S1~S5, sample data from different areas of the urban pipe network are collected. The sample data includes urban pipe network structural parameters (such as pipe inner diameter, pipe wall thickness, material code and topological coordinates), environmental status data, and differential entropy feature matrix after electrical variable compensation, dynamic gain adjustment, anti-interference processing and differential entropy feature extraction. ;

[0188] Environmental status data includes temperature and humidity;

[0189] The labeling is carried out in combination with the historical maintenance records of the urban pipe network and the evaluation of technical personnel. Specifically, the technical personnel mark the health status label of each sample based on the electrical variable data, on-site inspection reports (such as corrosion depth measurement, leakage point location) and urban pipe network operation logs; a labeled training data set is formed, in which the differential entropy feature matrix of each sample is Associated with the corresponding health status label for supervised learning of the health status evaluation model.

[0190] Health status labels include "normal state", "mild corrosion", "moderate corrosion", "severe corrosion" and "leakage"; normal state means that the pipeline has no signs of corrosion or leakage, mild corrosion means that the pipeline wall thickness loss is less than 20%, moderate corrosion means that the pipe wall thickness loss is 20%-50%, severe corrosion means that the pipe wall thickness loss is greater than 50%, and leakage means that there is fluid leakage or abnormal pressure in the pipeline.

[0191] S7. Build a health status assessment model using an end-to-end complex neural network architecture, train the health status assessment model, obtain a trained health status assessment model, preserve the amplitude-phase coupling relationship between features, and optimize the nonlinear response capability based on the pipeline network fault characteristics;

[0192] S701. Define the health status evaluation model structure;

[0193] The health status evaluation model adopts an end-to-end complex neural network architecture, with the input being the differential entropy feature matrix , the output is the probability distribution of the health status of the urban pipeline network.

[0194] The core of the model consists of four sequential modules: adaptive feature gating layer, dual-domain activation function layer, gradient-stabilized batch normalization layer, and fully connected classification layer.

[0195] The amplitude-phase coupling relationship between features is preserved through complex matrix operations, and the nonlinear response capability is optimized according to the fault characteristics of urban pipeline networks.

[0196] S702. Constructing an adaptive feature gating layer based on the adaptive feature gating mechanism;

[0197] S7021. Calculate feature gating weights;

[0198] The input feature vector is processed by combining two layers of trainable weight matrices and activation functions. The first layer uses the Sigmoid activation function to generate the basic weights. The second layer modulates the weight range through the hyperbolic tangent function. The outputs of the two layers are multiplied element by element to generate a gated weight vector. The gated weight vector represents the modulation strength of each feature dimension and is used to dynamically evaluate the importance of features at each scale.

[0199] Gating weight vector Belongs to the 3-dimensional real number space, and the element range is , the calculation formula is:

[0200] ,

[0201] Where, For the Sigmoid activation function, the output value range ; is the first weight matrix, belonging to Real matrix space, trainable parameters; is the input feature vector, the output from the differential entropy feature extraction step; is the bias vector, which belongs to the 3D real space and can be trained as a parameter; It is an element-by-element multiplication operation; is the hyperbolic tangent function, and the output range is ; is the second weight matrix, belonging to Real matrix space, trainable parameters.

[0202] S7022. Perform outer product feature enhancement;

[0203] The gate weight vector is combined with the input feature vector to generate a two-dimensional enhanced feature matrix. Real matrix space, which can amplify the contribution of key feature dimensions while preserving the original feature relationships;

[0204] Enhanced feature matrix The calculation formula is:

[0205] ,

[0206] Where, is the outer product operation, that is The operation of the term is defined as .

[0207] When deep neural networks process multi-scale features, conventional ReLU activation functions are insufficiently responsive to subtle fault signatures and cannot dynamically adjust the contribution weights of features at different scales, resulting in reduced fault sensitivity and inadequate feature utilization. The S702 learns feature importance through trainable parameters, generates a gated weight vector, and performs an outer product operation with the input feature vector. This eliminates the contribution of key feature dimensions amplified by power frequency interference, enhances fault-related features, and preserves the original feature relationships, improving the network's responsiveness to subtle faults.

[0208] S703. Constructing a dual-domain activation function layer based on the dual-domain activation function;

[0209] S7031. Calculate the eigenvalue matrix;

[0210] Perform a modulus operation on the enhanced feature matrix to extract the amplitude information of each element and generate a real-valued feature modulus matrix to represent the feature intensity information, which is expressed as:

[0211] ,

[0212] Where A represents the characteristic module value matrix, which belongs to Real matrix space ; It is a modulo operation.

[0213] S7032. Calculate the characteristic phase angle matrix;

[0214] Perform phase angle calculation on the enhanced feature matrix to extract the phase direction information of each element and obtain the feature phase angle matrix to represent the feature direction information;

[0215] The characteristic phase angle matrix belongs to Real matrix space, element range is , characteristic phase angle matrix Expressed as:

[0216] ,

[0217] Where, For phase angle operation, due to the enhanced characteristic matrix is a real matrix;

[0218] Enhanced feature matrix The phase angle of the element in row i and column j in The calculation formula is:

[0219] ,

[0220] Where, is the enhanced feature matrix The element in row i and column j;

[0221] S7033. Amplitude path nonlinear transformation;

[0222] The eigenvalue matrix is ​​input into the trainable weight matrix, and a nonlinear transformation is performed through the Sigmoid activation function. The transformation result is multiplied element by element with the original eigenvalue matrix to enhance the response capability of the abnormal amplitude feature, which is expressed as:

[0223] ,

[0224] Where, Represents the transformed eigenvalue matrix, belonging to Real matrix space; A is the characteristic module value matrix, which is Calculated, belongs to Real matrix; is the amplitude subnetwork weight matrix, belonging to Real matrix space, trainable parameters;

[0225] S7034.Phase path nonlinear transformation;

[0226] The phase angle matrix is ​​input into the trainable weight matrix, and a nonlinear transformation is performed through the hyperbolic tangent function. The transformation result is multiplied element by element with the original phase angle matrix to enhance the responsiveness of the phase distortion feature, which is expressed as:

[0227] ,

[0228] Where, is the transformation phase matrix, belonging to Real matrix space; is the phase subnetwork weight matrix, belonging to Real matrix space, trainable parameters;

[0229] S7035. Complex domain feature reconstruction;

[0230] The transformed eigenvalue matrix is ​​used as the real part and the transformed phase matrix is ​​used as the imaginary part to generate a complex eigenvalue matrix, thereby eliminating the power frequency interference and retaining the coupling relationship between amplitude and phase, which is expressed as:

[0231] ,

[0232] Where, Represents the output matrix of the dual-domain activation function, belonging to Complex matrix space, as the output of this step; The addition method in the terms is matrix addition operation, and the real part and the imaginary part are added separately; Is an imaginary unit.

[0233] Urban pipeline health status classification requires response to characteristic amplitude anomalies and phase distortion. Conventional activation functions only process real-valued characteristic amplitude information, ignoring complex-domain phase relationships. This ineffectively utilizes phase information, resulting in insufficient sensitivity to phase distortion characteristics of faults such as pipeline corrosion or leakage. S703 defines a dual-domain activation function in the complex domain. By processing the characteristic modulus and phase angle matrices in parallel, it performs nonlinear transformations and reconstructs complex features. This eliminates power frequency interference while preserving the coupling relationship between amplitude and phase, enhancing the response to amplitude anomalies and phase distortion, and improving fault detection accuracy.

[0234] S704. Construct a gradient-stabilized batch normalization layer based on gradient-stabilized batch normalization;

[0235] S7041. Separate the real and imaginary parts;

[0236] Decompose the input complex feature matrix into independent real and imaginary matrices, expressed as:

[0237] ,

[0238] ,

[0239] In the formula, R represents the real matrix, which belongs to Real matrix space; Represents the imaginary part matrix, belonging to Real matrix space.

[0240] S7042. Calculate real batch statistics;

[0241] Based on the current training batch data, calculate the mean and variance of all elements of the real matrix. The batch mean represents the trend of the data concentration, and the batch variance represents the degree of data dispersion, which can be expressed as:

[0242] ,

[0243] ,

[0244] Where, represents the real part batch mean, which is a real number; B is the training batch size; b is a positive integer; Represents the element in row i and column j of the real part matrix of the bth sample; Represents the real batch variance, which is a non-negative real number; The sum of the items represents the sum of the matrix elements of all samples in the batch.

[0245] S7043. Normalize the real part and scale and translate;

[0246] Use the real batch statistics to normalize the real matrix, restore the feature expression ability through trainable scaling parameters and translation parameters, and output the normalized real matrix;

[0247] Normalized real part matrix Expressed as:

[0248] ,

[0249] Where, is the real part scaling parameter, which is a trainable real number vector with an initial value of 1; To prevent the denominator from being 0, you can set =0.0001; is the real part translation parameter, which is a trainable real number vector and has an initial value of 0;

[0250] S7044. Calculate imaginary batch statistics;

[0251] Based on the current training batch data, the mean and variance of all elements of the imaginary matrix are calculated independently of the real part statistics to avoid interference from dimensional differences. It is expressed as:

[0252] ,

[0253] Where, represents the imaginary batch mean, which is a real number; Represents the element in row i and column j of the imaginary part matrix of the bth sample; Represents the imaginary batch variance, which is a non-negative real number;

[0254] S7045. Normalize the imaginary part and scale and translate;

[0255] Use the imaginary part batch statistics to normalize the imaginary part matrix, adjust the imaginary part feature distribution through independent trainable scaling parameters and translation parameters, and output the normalized imaginary part matrix;

[0256] Normalized imaginary part matrix Expressed as:

[0257] ,

[0258] Where, is the imaginary part scaling parameter, which is a trainable real number vector; is the imaginary translation parameter, which is a trainable real vector.

[0259] S7046. Combine the real and imaginary parts;

[0260] The normalized real and imaginary matrices are recombined into a complex characteristic matrix to eliminate power frequency interference, eliminate the dimension difference between the real and imaginary parts, and stabilize the gradient propagation process, which can be expressed as:

[0261] ,

[0262] Where, represents the decoupled batch normalization output matrix, which belongs to Complex matrix space; The complex number addition operation is performed by adding the real part and the imaginary part separately.

[0263] The real and imaginary parts of electrical variable features have dimensional differences. Conventional batch normalization layers normalize complex features as a whole, which can easily lead to gradient oscillations and reduce model training stability and convergence speed. The S704 uses a decoupled batch normalization mechanism. By separating the real and imaginary matrices, independently calculating batch statistics, normalizing them, and recombining them, it eliminates the impact of power frequency interference and dimensional differences, stabilizes the gradient propagation process, and improves model training efficiency and stability.

[0264] S705. Output the classification result through the fully connected classification layer;

[0265] Decoupled Batch Normalization Output Matrix Input to the fully connected classification layer and output the C-dimensional raw score vector, which represents the model's original prediction score for the category; use the Softmax function to convert the raw score into a probability distribution to obtain the prediction probability of each category; define The model predicts that the sample belongs to The probability of each category is used to calculate the fuzzy boundary focus loss, and the category with the largest probability is taken as the predicted category for this iteration.

[0266] S706. Calculate blur boundary focus loss;

[0267] S7061. Calculate fuzzy membership;

[0268] Based on the Euclidean distance between the true label of the sample and the center of the pre-clustering class, combined with the variance of the intra-class distribution, the fuzzy membership of the sample to each category is calculated;

[0269] The sample belongs to The fuzzy membership of categories is real number, and The larger the value, the higher the membership degree, and the sample belongs to Fuzzy membership of categories The calculation formula is:

[0270] ,

[0271] Where, is the true label vector of the sample; For the The feature cluster center vector of each category is obtained by differential entropy feature matrix The feature pre-clustering is obtained, and the differential entropy feature matrix of the training set can be set by the K-means clustering algorithm Perform pre-clustering, and the cluster center of each category is obtained by minimizing the intra-class distance; is the L2 norm, the same as the Euclidean distance operation method, The term represents the difference between the sample label and the class center; For the The feature distribution variance of each category is calculated by calculating the sample within the category to the Feature cluster center vector The Euclidean distance variance is obtained and used to adjust the sensitivity of the fuzzy membership; Is a positive integer.

[0272] S7062. Calculate class weight factors;

[0273] According to the frequency of occurrence of each category in the training set, the inverse frequency square root weight is calculated to alleviate the category imbalance problem and improve the recognition sensitivity of rare states.

[0274] No. The weight factors of the categories are positive real numbers, and ;No. Weight factor for each category The calculation formula is:

[0275] ,

[0276] Where, For the The frequency of occurrence of each category in the training set;

[0277] The calculation formula is:

[0278] ,

[0279] Where, For the The number of samples in each category; is the total number of samples.

[0280] S7063. Calculate the focus modulation factor;

[0281] A modulation factor is constructed based on the model prediction probability. The lower the prediction probability, the larger the factor value is, which strengthens the weight of low prediction probability samples in the loss calculation, expressed as:

[0282] ,

[0283] Where, represents the focal modulation factor, which is a real number, and ; The model predicts that the sample belongs to The probability of the categories, and , output by the Softmax function; is the focusing factor, which is a positive integer and controls the weighted intensity of samples with low prediction probability. The larger the focusing factor, the more attention is paid to difficult samples. .

[0284] S7064. Calculate weighted loss components;

[0285] Combining fuzzy membership, class weight factor and focus modulation factor, the cross entropy loss of each category is weighted. The higher the membership, the rarer the category and the more difficult the sample, the greater the loss weight, which is expressed as:

[0286] ,

[0287] Where, Indicates the The weighted loss component of each category, which is a real number.

[0288] S7065. Sum to get the total loss;

[0289] The weighted loss components of all categories are summed to generate the fuzzy boundary focus loss, which is used to optimize the gradient update of boundary samples and improve the generalization ability of the model in the state transition area. It is expressed as:

[0290] ,

[0291] Where, represents the fuzzy boundary focusing loss, which is a non-negative real number, and , as the objective function of model training; is the total number of categories.

[0292] There is ambiguity in the transition zone of the health status of urban pipe networks. Conventional cross entropy loss over-punishes boundary samples, which easily leads to a decrease in the generalization ability of the model in the state transition zone and fails to accurately reflect the continuous changes in the health status. S706 is based on the fuzzy membership function. By calculating the fuzzy membership of the sample, the class weight factor and the focus modulation factor, the weighted cross entropy loss is added and summed, thereby eliminating the power frequency interference and optimizing the gradient update of the boundary samples, improving the generalization ability of the model in the state transition zone, and alleviating the class imbalance problem. The recognition ability of different models for five types of urban pipe network health status is compared, including support vector machines, random forests, conventional convolutional neural networks and S706 complex neural network models. Figure 6 The accuracy bar chart shows that the vertical axis performance is measured by accuracy, and the classification accuracy of S706 is significantly ahead. Figure 7 The confusion matrix heat map reveals that the S706's recognition accuracy for high-risk states such as "severe corrosion" and "leakage" is over 93%, and the misjudgment rate of adjacent states (such as mild-moderate corrosion) is controlled within 6%, proving that the dual-domain activation function and decoupled batch normalization mechanism successfully retain the amplitude-phase coupling relationship of fault characteristics.

[0293] S707. Repeat iterations S702 to S705 until the preset stop iteration condition is met;

[0294] The model is iteratively trained using mini-batch gradient descent. Each iteration randomly samples a batch of samples from the training set, executes steps S702-S705 sequentially to perform forward propagation, and calculates the predicted probability. All model trainable parameters are updated by calculating the fuzzy boundary focus loss and backpropagating the gradients. During training, validation set loss and classification accuracy are continuously monitored. Iterations are terminated when validation loss fails to decrease for 10 consecutive rounds or the total number of training rounds reaches a preset upper limit. The model parameters with the best validation set performance are saved, with a preset upper limit of 2000.

[0295] S8. After executing S1 to S5 in the urban pipe network area to be tested, obtain the test data, load the trained health status evaluation model, input the test data into the trained health status evaluation model for forward reasoning, and the trained health status evaluation model outputs the probability of each health status label, and takes the category with the highest probability as the evaluation result.

Claims

1. An artificial intelligence-based urban pipe network health status evaluation method, characterized by: include: S1. Obtain urban pipe network structural parameters and measure electrical variable data; S2. Synchronously acquire the working electrode voltage and the reference electrode voltage based on the four-electrode differential measurement method; Utilize dynamic compensation mechanism to output compensation voltage value; S3. Calculate the signal fluctuation intensity and dynamically adjust the gain coefficient, and calculate the adjusted signal based on the gain coefficient; S4. synchronously collect the positive-phase and negative-phase signals, perform differential operations, apply a frequency-domain Gaussian mask to suppress noise, and obtain an anti-interference output vector; S5. Construct an analytical signal through Hilbert transform and multi-scale wavelet decomposition, calculate the energy probability distribution and differential entropy characteristics, and build a differential entropy feature matrix; The steps of S5 include: S501. Performing a Hilbert transform on the input anti-interference output signal to generate a complex analytical signal containing amplitude and phase information; S502. Perform multi-scale decomposition of the complex analytic signal using the Morlet wavelet basis function, and extract wavelet components of different frequency bands at preset high, medium, and low frequency scales; S503. For all wavelet components at each scale, calculate the ratio of the energy of each component to the total energy of the scale to form an energy probability distribution: S504. Calculate differential entropy features based on the energy probability distribution of each scale; S505. Integrate the differential entropy features of the three preset scales of high frequency, medium frequency and low frequency to form a differential entropy feature matrix; S6. Build and label the training dataset; S7. constructing a health status evaluation model using an end-to-end complex neural network architecture, training the health status evaluation model, and obtaining a trained health status evaluation model; S8. After executing S1 to S5 in the urban pipe network test area, obtain the test data, input the test data into the trained health status evaluation model, and the trained health status evaluation model outputs the probability of each health status label, and takes the category with the highest probability as the evaluation result.

2. The urban pipe network health status evaluation method based on artificial intelligence according to claim 1 is characterized in that: The steps of S1 include: S101. Obtaining urban pipe network structural parameters: S102. Synchronously measure electrical variable data; Install four-electrode sensing units at key nodes of the urban pipe network to achieve high-precision acquisition of electrical variable data; Electrical variable data include: conductivity raw signal, current and voltage related variables; current and voltage related variables include excitation current, working electrode voltage vector, reference electrode voltage vector; Key nodes in urban pipeline networks include flanges, elbows, and valves.

3. The urban pipe network health status evaluation method based on artificial intelligence according to claim 1 is characterized in that: The steps of S2 include: S201. Based on the four-electrode differential measurement principle, a constant current is injected while the working electrode voltage and the reference electrode voltage are simultaneously acquired. A dynamic compensation mechanism is used to eliminate baseline drift caused by the electrode-tube wall interface impedance using a calibration resistor and reference loop impedance, and the compensated voltage value is output. S202. Under no-load conditions without connection to the urban pipeline network, inject a calibration current and measure the no-load reference voltage. The calibration resistance is calculated by the ratio of the calibration current to the no-load reference voltage and used as a constant in the dynamic compensation formula.

4. The urban pipe network health status evaluation method based on artificial intelligence according to claim 1 is characterized in that: The steps of S3 include: S301. Taking the current sampling point as the center, select the historical compensation voltage data points contained in the sliding window, and calculate the average value of the selected historical compensation voltage data points as the local voltage reference value; S302. Calculate the standard deviation of the compensation voltage data within the sliding window, combine the range scaling factor and the anti-zero constant, and determine the adaptive gain coefficient; S303. Multiply the gain coefficient of the current sampling point by the compensation voltage value, and output the adjusted signal.

5. The method for evaluating the health status of urban pipe networks based on artificial intelligence according to claim 1 is characterized in that: The steps of S4 include: S401. In a single excitation cycle, synchronously collect the positive phase voltage signal and the negative phase voltage signal to form a signal pair for subsequent differential processing; S402. Perform differential operation on the positive and negative phase signals within the same excitation cycle and take the absolute value of the result; S403. Averaging the differential absolute value signals of multiple consecutive excitation cycles to reduce the influence of random noise, converting the averaged signal into the frequency domain through fast Fourier transform to obtain a frequency domain representation; S404. Multiply the frequency domain Gaussian mask and the average frequency domain signal element by element to obtain a filtered frequency domain signal; S405. Perform inverse Fourier transform on the filtered frequency domain signal, convert it back to the time domain, and output an anti-interference output vector.

6. The urban pipe network health status evaluation method based on artificial intelligence according to claim 1 is characterized in that: The steps of S7 include: S701. Define the health status evaluation model structure; The health status evaluation model adopts an end-to-end complex neural network architecture, with the input being the differential entropy feature matrix , the output is the probability distribution of the health status of the urban pipeline network; S702. Constructing an adaptive feature gating layer based on the adaptive feature gating mechanism; S703. Constructing a dual-domain activation function layer based on the dual-domain activation function; S704. Construct a gradient-stabilized batch normalization layer based on gradient-stabilized batch normalization; S705. Output the classification result through the fully connected classification layer; S706. Calculate blur boundary focus loss; S707. Repeat iterations S702 to S705 until a preset stop iteration condition is met.

7. The method for evaluating the health status of urban pipe networks based on artificial intelligence according to claim 6 is characterized in that: The step of S703 includes: S7031. Perform a modulus operation on the enhanced feature matrix to extract the amplitude information of each element and generate a feature modulus matrix; S7032. Perform phase angle calculation on the enhanced feature matrix to extract the phase direction information of each element to obtain a feature phase angle matrix; S7033. Input the eigenvalue matrix into the trainable weight matrix, perform nonlinear transformation through the Sigmoid activation function, and multiply the transformation result element-by-element with the original eigenvalue matrix; S7034. Input the phase angle matrix into the trainable weight matrix, perform nonlinear transformation using the hyperbolic tangent function, and multiply the transformation result element-by-element with the original phase angle matrix; S7035. Use the transformed eigenvalue matrix as the real part and the transformed phase matrix as the imaginary part to generate a complex eigenmatrix.

8. The method for evaluating the health status of urban pipe networks based on artificial intelligence according to claim 6 is characterized in that: The steps of S704 include: S7041. Decompose the input complex feature matrix into independent real and imaginary matrices; S7042. Based on the current training batch data, calculate the mean and variance of all elements of the real matrix; S7043. Use the real batch statistics to normalize the real matrix, restore the feature expression ability through the trainable scaling parameters and translation parameters, and output the normalized real matrix; S7044. Based on the current training batch data, calculate the mean and variance of all elements of the imaginary matrix; S7045. Normalize the imaginary part matrix using the imaginary part batch statistic, adjust the imaginary part feature distribution through independent trainable scaling parameters and translation parameters, and output the normalized imaginary part matrix; S7046. Recombine the normalized real matrix and imaginary matrix into a complex characteristic matrix.

9. The method for evaluating the health status of urban pipe networks based on artificial intelligence according to claim 6 is characterized in that: The steps of S706 include: S7061. Calculate the fuzzy membership of the sample to each category based on the Euclidean distance between the true sample label and the pre-clustered cluster center and the variance of the intra-class distribution; S7062. Calculate the inverse frequency square root weight based on the frequency of occurrence of each category in the training set; S7063. Construct modulation factors based on model prediction probabilities; S7064. Combining fuzzy membership, class weight factor and focus modulation factor to weight the cross entropy loss of each class; S7065. Sum the weighted loss components of all categories to generate the fuzzy boundary focus loss.

Images