Urban pipe network operation state identification method

Abstract

The invention relates to the field of electrical variable measurement, in particular to an urban pipe network operation state identification method, which comprises the steps of collecting and marking multiple groups of urban pipe network operation state data, preprocessing the multiple groups of marked urban pipe network operation state data, and performing feature extraction through a multi-scale feature pyramid; constructing a deep gating adaptive network; initializing the weight of the depth gating adaptive network; calculating a multi-target joint total loss function of the deep gating adaptive network; iterative training and parameter updating of the deep gating adaptive network are carried out; acquiring new urban pipe network operation state data, and processing the new urban pipe network operation state data to obtain a state label corresponding to the new urban pipe network operation state data; an existing urban pipe network operation state identification method has the problem that the urban pipe network operation state identification effect is poor. The urban pipe network operation state identification method provided by the invention has a good urban pipe network operation state identification effect.

Description

Technical Field

[0001] The present invention relates to the field of measuring electrical variables, and particularly to a method for identifying the operating state of urban pipe networks. Background Art

[0002] Urban pipe networks, also known as urban underground pipelines, refer to various pipelines and their ancillary facilities buried underground in cities, and are important infrastructure and "lifelines" for ensuring urban operation. Urban pipe networks can be classified into urban water supply pipe networks, urban drainage pipe networks, urban gas pipe networks, urban central heating pipe networks, etc. according to their uses. As the core system for the operation of urban infrastructure, the operating state of urban pipe networks is directly related to the stability and safety of public services such as urban water supply, heating, and drainage.

[0003] With the acceleration of the urbanization process, the structure of urban pipe networks has become increasingly complex. The traditional manual inspection method has been difficult to meet the needs of modern urban management. People have begun to use urban pipe network monitoring equipment to extract electrical variable data at key nodes of urban pipe networks and combine big data model analysis to monitor the operating state of urban pipe networks. The existing methods for identifying the operating state of urban pipe networks have the following problems when identifying the operating state of urban pipe networks: (1) Existing identification methods usually use fixed-window statistics or single-scale frequency-domain analysis for feature extraction, and cannot capture the coupling relationship between sudden transients and long-term trends in electrical variable data, that is, they cannot capture micro mutations and macro trends at the same time, and are prone to missing key state transition signals, thus affecting the training effect of subsequent big data models and the identification effect; (2) Existing identification methods use deep neural networks to identify the state of electrical variable data, but the dynamic correlation modeling of the time series features of deep neural networks is insufficient and the fixed activation function is difficult to adapt to multi-condition changes. When facing different operating states of urban pipe networks, it is difficult to dynamically adjust the modeling strategy, and there is a lack of a feature retention and selection mechanism for weak features and complex background signals, resulting in easy drowning of fault features, making the existing identification methods have poor identification effects on boundary samples and complex working conditions; and the existing identification methods only use cross-entropy loss for the model output, do not fully pay attention to difficult-to-classify samples, and lack optimization of the separability of the feature space, resulting in fuzzy classification boundaries and easy confusion between faults and normal states; Therefore, the existing methods for identifying the operating state of urban pipe networks have the problem of poor identification effect on the operating state of urban pipe networks. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the prior art and provide a method for identifying the operating state of urban pipe networks with better identification effect on the operating state of urban pipe networks.

[0005] To solve the above technical problem, a method for identifying the operating state of urban pipe networks provided by the present invention includes: S1. Collect multiple groups of urban pipe network operation status data and perform status annotation on each group of urban pipe network operation status data to obtain multiple groups of annotated urban pipe network operation status data. The multiple groups of annotated urban pipe network operation status data together constitute a training dataset; The urban pipe network operation status data includes the measured values of electrical variables at multiple key nodes in the urban pipe network; S2. Preprocess the multiple groups of annotated urban pipe network operation status data; S3. Extract features through a multi-scale feature pyramid; S4. Construct a deep gated adaptive network; Perform soft selection on the input features through a gating mechanism, and retain key information by combining memory vectors; mix the ReLU and hyperbolic tangent activation functions in the residual block, adjust the nonlinear intensity through learnable parameters, and perform channel-level scaling on the original input during skip connection; use a dynamic convolution kernel to process time series data; S5. Initialize the weights of the deep gated adaptive network; S6. Calculate the multi-objective joint total loss function of the deep gated adaptive network; The total loss function consists of a cross-entropy loss for focusing on correct classification, a focal loss for dynamically adjusting the sample weights through predicted probabilities, and a maximum mean discrepancy loss for aligning the distributions of different classes in the feature space; S7. Iteratively train and update the parameters of the deep gated adaptive network to obtain a trained deep gated adaptive network; S8. Collect new urban pipe network operation status data, preprocess the new urban pipe network operation status data to obtain preprocessed new urban pipe network operation status data, input the preprocessed new urban pipe network operation status data into the trained deep gated adaptive network, and the trained deep gated adaptive network analyzes the preprocessed new urban pipe network operation status data and outputs the status label corresponding to the new urban pipe network operation status data.

[0006] As a further improvement of the present invention: The method for obtaining the measured values of electrical variables at multiple key nodes in the urban pipe network in S1 includes: S101. Deploy voltage-current composite sensors at key nodes; S102. The voltage-current composite sensor discretely samples the sensed analog electrical signal through an analog-to-digital conversion module to obtain a sampled electrical signal, and the sampling frequency is dynamically adjusted according to the working conditions, including: S1021. Input the analog electrical signal; S1022. Condition the front-end signal to obtain a processed analog signal; S1023. Input the processed analog signal into the analog-to-digital conversion module, and the analog-to-digital conversion module internally converts the continuous analog waveform into 24-bit discrete digital values; S103. The sampled electrical signals are channel numbered according to their corresponding measurement nodes.

[0007] As a further improvement of the present invention: The status labels in S1 include "fault" and "normal operation".

[0008] As a further improvement of the present invention: The preprocessing in S2 includes: S201. Perform normalization processing on multiple groups of labeled urban pipe network operation status data to obtain normalized labeled urban pipe network operation status data; S202. Perform noise reduction processing on the normalized labeled urban pipe network operation status data.

[0009] As a further improvement of the present invention: The feature extraction by the multi-scale feature pyramid in S3 includes: S301. Perform empirical mode decomposition on the noise-reduced signal, extract the modal components reflecting transient oscillation characteristics, calculate the instantaneous energy of each component through Hilbert transform, and extract multiple statistics to form a transient feature vector; S302. Use the dynamic time warping algorithm to align all sample signals with the reference mean sequence, extract multiple trend features from the optimal matching path, and perform wavelet packet decomposition; S303. Select the frequency band node with the highest energy entropy ratio to reconstruct the frequency domain signal, extract the frequency domain features, and splice the three types of features to form a multi-scale pyramid.

[0010] Preferably, S301 performs empirical mode decomposition on the noise-reduced signal, extracts the modal components reflecting transient oscillation characteristics, calculates the instantaneous energy of each component through Hilbert transform, and extracts multiple statistics to form a transient feature vector, including: S3011. The measured value of the i-th channel of the electrical variable at time t after wavelet noise reduction processing Perform empirical mode decomposition to obtain modal components; S3012. Calculate the Hilbert marginal spectrum energy of each modal component; S3013. Construct a transient feature vector.

[0011] Preferably, S303 selects the frequency band node with the highest energy entropy ratio to reconstruct the frequency domain signal, extracts the frequency domain features, splices the three types of features to form a multi-scale pyramid, and realizes the full-scale characterization from microscopic transients to macroscopic trends, including: S3031. Perform multi-layer wavelet packet decomposition on the signal, calculate the weighted sum of the proportion of the energy of each node in the total amount and its frequency band information entropy as the node weight coefficient, select the three nodes with the highest weights, reconstruct the frequency domain components using their basis functions and weight coefficients, extract relevant features, and splice the transient, trend, and frequency domain features into a multi-scale feature pyramid; S3032. Select the nodes with the top three energy entropy ratios to form frequency domain features and match the frequency domain sparsity of electrical variables.

[0012] As a further improvement of the present invention: The initialization of the weights of the deep gated adaptive network in S5 includes: S501. Calculate the quartiles of each dimension of the multi-scale time series features extracted by the multi-scale feature pyramid; S502. Construct an initialization weight matrix whose elements satisfy a uniform distribution.

[0013] As a further improvement of the present invention: The total loss function in S6 has the following calculation formula: , wherein, is the weight of the cross-entropy loss; is the cross-entropy loss; is the focal loss; is the weight coefficient of the maximum mean discrepancy loss; is the maximum mean discrepancy loss.

[0014] As a further improvement of the present invention: The iterative training and parameter update of the deep gated adaptive network in S7 include: S701. Define the parameter set of the model as ; S702. In each iteration process, calculate the gradient of the parameter with respect to the total loss function , and update the model parameters; S703. Combine the momentum mechanism, and the momentum term provides inertia for parameter update by accumulating historical gradient information.

[0015] The beneficial effects of the present invention are as follows: A method for identifying the operation status of urban pipe networks provided by the present invention has a good effect on identifying the operation status of urban pipe networks.

[0016] This method comprehensively captures local mutations and global evolution behaviors in the operation of urban pipe networks by constructing multi-scale features. It uses a gating mechanism and memory vectors to dynamically adjust feature retention. The convolutional kernel is dynamically adjusted by self-attention to enhance the modeling ability for temporal dependencies. Cross-entropy, focal loss, and maximum mean discrepancy loss are used to balance classification accuracy, attention to difficult examples, and separability in the feature space distribution, improving the model's recognition effect on boundary samples and complex working conditions, thus making this method have a good recognition effect on the operation status of urban pipe networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is the principle block diagram of the present invention; Figure 2 is the data processing flow chart in the present invention; Figure 3 is the comparison diagram of the effects of different noise reduction methods in the present invention; Figure 4 is the comparison diagram of the method robustness under different noise levels in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0018] The following further elaborates in detail the specific embodiments of the present invention with reference to the accompanying drawings.

[0019] As Figure 1 , Figure 2 shown, a method for identifying the operation status of urban pipe networks provided by the present invention includes: S1. Collect multiple groups of urban pipe network operation status data, and let those skilled in the art manually identify each group of urban pipe network operation status data and perform status annotation on each group of urban pipe network operation status data to obtain multiple groups of annotated urban pipe network operation status data. The multiple groups of annotated urban pipe network operation status data together constitute a training data set; The urban pipe network operation status data includes the measured values of electrical variables at multiple key nodes in the urban pipe network. Denote the measured value of the i-th channel of the electrical variable at time t as ; The method for obtaining the measured values of electrical variables at multiple key nodes in the urban pipe network is as follows: Set sensors at multiple key nodes in the urban pipe network to obtain the current changes, voltage changes, etc. (i.e., the measured values of electrical variables) of each key node (pumping station, valve controller, or motor). The current change includes the current value directly measured by the current sensor and its fluctuations; The method for obtaining the measured values of electrical variables at multiple key nodes in the urban pipe network specifically includes: S101. Deploy high-sensitivity voltage-current composite sensors at key nodes such as pumping stations, valve nodes, frequency converter control units, and pressure regulation points. The voltage-current composite sensors have a wide dynamic response range and can cover low-frequency and high-frequency changes during the operation of the municipal pipe network; S102. The voltage-current composite sensor performs discrete sampling on the sensed analog electrical signal through the analog-to-digital conversion module to obtain the sampled electrical signal. The sampling frequency is dynamically adjusted according to the working conditions. For example, the sampling frequency is set to 1000 Hz to ensure sufficient transient information is retained in case of emergencies. The sensor performs discrete sampling on the sensed analog electrical signal through the analog-to-digital conversion module. For the current monitoring points of urban pipe network pump stations, a 24-bit high-precision analog-to-digital conversion module is used. The discrete sampling of the sensed analog electrical signal through the analog-to-digital conversion module includes: S1021. Input the analog electrical signal; The Hall current sensor installed on the electric control device of the pump station inlet valve outputs an analog current signal that reflects the working state of the pump station in real time , with a voltage range of 0 - 5V and a bandwidth of 0 - 500 Hz; S1022. Condition the front-end signal to obtain the processed analog signal; To ensure the input signal quality of the analog-to-digital conversion module, the analog signal first passes through a low-pass filter and a programmable gain amplifier to filter out high-frequency interference and amplify the signal to the full-scale input range, obtaining the processed analog signal. The cut-off frequency of the low-pass filter is set to 600 Hz; S1023. Input the processed analog signal into the analog-to-digital conversion module and set the sampling rate = 1000 Hz. The analog-to-digital conversion module internally converts the continuous analog waveform into 24-bit discrete digital values; The measured value of the i-th channel of the electrical variable at time t The calculation formula is: , In the formula, represents the function mapping process of the analog-to-digital conversion module for discretely sampling the input analog electrical signal; is the analog current signal; S103. The sampled electrical signal is channel numbered according to its corresponding measurement node. For example, the measured value of the i-th channel of the electrical variable at time t is expressed as .

[0020] By deploying electrical signal sensors at key nodes of the urban pipe network, physical quantities such as voltage and current during operation are collected in real time and converted into standardized electrical signal data to achieve real-time monitoring of the operation status of the urban pipe network. The electrical variable data serves as the input basis for subsequent state recognition modeling. Each data stream collected is attached with meta-information such as timestamp, node number, and collection device number to achieve consistent synchronization of multi-source data across devices, ensuring the consistency of subsequent sliding window normalization and multi-scale analysis. To meet the needs of model supervised training, the electrical variable data at each time step is labeled according to the operation records of the pipe network scheduling system, and status labels are set. The status labels are obtained from operation and maintenance records, abnormal alarm logs, or expert manual judgment, and a labeled data set for training is constructed.

[0021] The key nodes of the urban pipe network are pump stations, valve controllers, motors, etc. set in the urban pipe network; Electrical variables refer to electrical parameters that reflect the operation status of electrical systems or equipment, which can usually be directly measured or indirectly calculated through sensors. Electrical variables include voltage, current, etc.; The status labels for status annotation include "fault" and "normal operation"; S2. Preprocess multiple groups of labeled operation status data of the urban pipe network; S201. Perform normalization processing on multiple groups of labeled operation status data of the urban pipe network to obtain normalized labeled operation status data of the urban pipe network; The measured values of the original electrical variable data at each moment t are processed according to different channels and time steps, including: using the data of the previous L time steps as a sliding window, calculating the arithmetic mean of the data within the window as the local mean, and then calculating the standard deviation of each data point within the window from this mean. After subtracting the current moment's measured value from the local mean, divide it by the sum of the standard deviation and a very small constant to obtain the normalization result; The measured value of the i-th channel of the normalized electrical variable at time t The calculation formula is: , In the formula, is the measured value of the i-th channel of the electrical variable at time t; is the mean within the sliding window; is the standard deviation of the sliding window; is a constant to prevent the divisor from being zero, ; During the calculation of , for the non-stationarity of electrical variables of the same attribute in the urban pipe network, such as sudden load changes, a sliding window is used to calculate local statistics to avoid distribution deviation caused by outliers in global normalization; The mean within the sliding window The calculation formula is: , where L is the window length; t represents the time step at the current moment; represents the historical time steps within the sliding window; is the measured value of the i-th channel of the original electrical variable at moment; The calculation formula for the window length L is: , where represents rounding up; T is the total time step length; is the logarithmic function with a default base of 10; The window length can adapt to the data scale to ensure that the window can cover sufficient historical information without generating too much noise due to being too long; The standard deviation of the sliding window The calculation formula is: , where L is the window length; t represents the time step at the current moment; represents the historical time steps within the sliding window; is the measured value of the i-th channel of the original electrical variable at moment; is the mean value within the sliding window; S202. Denoise the normalized and labeled urban pipe network operation status data to obtain the preprocessed and labeled urban pipe network operation status data; By performing multi-scale wavelet decomposition on the normalized signal, the characteristic coefficients of each decomposition layer are obtained. For each layer of coefficients, first calculate the median of its absolute value, convert it to the noise threshold of this layer through a preset conversion factor, and according to the proportional relationship between the absolute value of the coefficient and the threshold, adopt a segmented processing strategy, that is, the coefficients exceeding the threshold are retained according to the sign and the amplitude is reduced, the coefficients in the middle interval are attenuated proportionally, and the coefficients below the threshold are directly set to zero. Finally, the signal is reconstructed with the processed coefficients. The specific steps are as follows: S2021. Perform multi-scale wavelet decomposition on to obtain the eigenvalues after wavelet decomposition, and define as the k-th eigenvalue of the j-th wavelet decomposition layer; S2022. Dynamically calculate the threshold of each layer of wavelet decomposition; The threshold of the j-th wavelet decomposition layer of wavelet decomposition is calculated as: , where is the median symbol. Since the median will not be skewed in the presence of outliers and can better reflect the typical amplitude of background noise; is the k-th eigenvalue of the j-th wavelet decomposition layer; is the natural logarithm function; is the number of coefficients in the j-th layer, that is, the number of eigenvalues of the j-th wavelet decomposition layer; can adapt to non-Gaussian noise in electrical variables, such as the mixed distribution of sensor noise and true signals; The reason for setting the denominator of to 0.6745 is to make an unbiased estimate of the standard deviation of Gaussian white noise. When assuming that the noise follows a normal distribution, the conversion factor between the median absolute deviation and the standard deviation is .

[0022] S2023. Define the threshold function as a mixed form of piecewise hard and soft thresholds; The k-th eigenvalue of the j-th wavelet decomposition layer after threshold processing The calculation formula is: , In the formula, is the sign function; is the first threshold adjustment parameter, which can be set to ; is the second threshold adjustment parameter, which can be set to .

[0023] Regarding the mixed characteristics of the sudden transients and continuous noise of electrical variables, to achieve faithful noise reduction, in the piecewise processing strategy: For strong signals, that is, the screening condition is , retain the sign and reduce the amplitude through the first threshold adjustment parameter to avoid excessive attenuation of sudden fault signals; For medium signals, that is, the screening condition is , attenuate proportionally through the second threshold adjustment parameter to retain weak but potentially important transitional state features; For weak signals, directly set to zero to suppress noise.

[0024] S2024. By dynamically adjusting the threshold interval and different attenuation coefficients, both retain weak features and suppress noise; After reconstructing the signal, we get , which represents the measured value of the i-th channel of the electrical variable at time t after wavelet noise reduction processing.

[0025] Due to the characteristics of high noise, multi-scale fluctuations, and local outliers in electrical variable data, the existing technology uses conventional normalization methods such as global maximum or standard deviation to scale electrical variables. Therefore, it cannot cope with the distribution shift (time period distribution change) caused by the difference in noise levels (sudden disturbances) and sudden outliers in different time periods. Using local statistics to replace global normalization can effectively address the non-stationarity of electrical variables; wavelet denoising usually adopts a fixed threshold solution, but it is difficult to balance denoising and feature retention in the face of complex working conditions. Therefore, this technology uses the wavelet decomposition level to dynamically adjust the threshold and combines a segmented soft-hard hybrid strategy to effectively suppress non-Gaussian noise and sudden interference. As Figure 3 shown, by comparing the effects of different denoising methods on noisy electrical variables, the superiority of the adaptive denoising technology of this technology is verified. The experiment compared the complex situations of periodic fluctuations, sudden fault pulses, random noise, and sudden strong noise in real electrical variables. The experimental data showed the original noisy signal, the real signal, and the processing results of the conventional moving average, conventional wavelet denoising, and this technology. Among them, the conventional moving average method smoothed the noise but seriously blurred the fault characteristics. The conventional wavelet denoising method is effective for stationary noise but insufficient for sudden noise. This technology effectively suppresses various noise interferences while retaining the important fault pulse characteristics. Especially in the time period of the marked fault area, this technology accurately restored the shape and amplitude of the fault pulse, while the comparison methods either over-smoothed or left obvious noise, indicating the advantages of the combination of sliding window local statistics and multi-scale adaptive threshold processing in this invention, which can effectively distinguish useful signals from noise.

[0026] S3. Feature extraction is performed through a multi-scale feature pyramid; Construct hierarchical multi-scale features by fusing time-frequency analysis and statistical methods; Perform empirical mode decomposition on the denoised signal, extract the modal components reflecting transient oscillation characteristics, calculate the instantaneous energy of each component through Hilbert transform, and statistically calculate indicators such as the maximum value, standard deviation, etc. to form a transient feature vector; use the dynamic time warping algorithm to align all sample signals with the reference mean sequence, extract trend features such as cumulative distance and slope from the optimal matching path, and perform wavelet packet decomposition; select the frequency band node with the highest energy entropy ratio to reconstruct the frequency domain signal, extract the frequency domain features, splice the three types of features to form a multi-scale pyramid, and achieve a full-scale representation from micro-transients to macro-trends; S301. Transient feature extraction layer; Perform empirical mode decomposition on the denoised signal to obtain several modal components reflecting different oscillation characteristics. For each modal component, calculate the instantaneous phase derivative after its Hilbert transform, calculate the instantaneous energy, and select the energy sequences of the first five modal components, and extract five statistics of the maximum value, standard deviation, mean, skewness, and kurtosis respectively to form a feature vector; The measured value of the i-th channel of the electrical variable at time t after wavelet denoising processing is subjected to empirical mode decomposition to obtain mode components . The m-th mode component reflecting the local oscillation characteristics of the signal after decomposition is expressed as ; S3012. Calculate the Hilbert marginal spectrum energy of each mode component; The instantaneous energy of the m-th mode at time t is calculated by the formula: , where is the Hilbert transform result of; denotes differentiation with respect to time; is the arctangent function; can extract the instantaneous energy through Hilbert transform to capture short-term oscillations in the urban pipe network, such as pump start / stop and valve actions; S3013. Construct transient feature vectors; The transient feature vector is calculated by the formula: , where represents the maximum value of the spectral energy of the first mode; represents the standard deviation of the spectral energy of the second mode; represents the mean value of the spectral energy of the third mode; represents the skewness of the spectral energy of the fourth mode; represents the kurtosis of the spectral energy of the fifth mode; altogether contains 5 statistics; S302. Trend feature extraction layer; Adopt the dynamic time warping method to align the long-term trends of different samples, calculate the average value of the denoised signals of all samples at each time step as the reference sequence, for each sample signal, find the optimal time alignment path with the reference sequence, which minimizes the sum of the squared distances of the corresponding points of the two signals, and calculate the cumulative distance, path slope, and count the number of extreme points of the original signal according to the alignment path. The three constitute the trend features; S3021. Represent the wavelet denoising mean reference sequence of all samples at time t as , and the wavelet denoising mean reference sequence of all samples at time t is calculated by the formula: , where N is the number of samples; S3022. Calculate the dynamic time warping path for each sample; The time matching path after aligning the i-th sample to the reference path The calculation formula is: , In the formula, denotes finding the path with the minimum loss among all paths ; denotes the set of all possible time alignment paths; denotes the time step of the i-th sample; q denotes the time step of the reference sequence; denotes the wavelet denoising signal value of the i-th sample at the time step; denotes the mean value of the reference sequence at the q time step; It can align each sample with the reference sequence and solve the problem of time asynchrony in urban pipe network data, such as different node response delays; S3023. Extract trend features; The trend feature vector The calculation formula is: , In the formula, is the cumulative distance value on the optimal alignment path, and the cumulative distance of all matching points on the optimal path is calculated through dynamic programming; is the average slope of the alignment path, and the average slope is obtained by linearly fitting the path points ; is the number of peak values of the wavelet denoising signal, and it is obtained by detecting the number of local extreme points of ; S303. Frequency domain feature fusion layer; S3031. Perform multi-layer wavelet packet decomposition on the signal, calculate the weighted sum of the proportion of the energy of each node in the total energy and its frequency band information entropy as the node weight coefficient, select the three nodes with the highest weights, reconstruct the frequency domain components with their basis functions and weight coefficients, extract relevant features, and splice the transient, trend, and frequency domain features into a multi-scale feature pyramid; The wavelet packet reconstructed signal at time t The calculation formula is: , In the formula, n is a positive integer; is the decomposition depth, which can be set ; is the weighted coefficient of the n-th node; is the n-th wavelet packet basis function; The weighted coefficient of the n-th node The calculation formula is as follows: , wherein, is the energy of the nth node, representing the energy intensity of the signal in this frequency band; is the total energy, representing the total energy of all nodes; is the entropy weight coefficient, which can be set to = 0.2; is the information entropy of the frequency distribution of the nth node; S3032. Select the nodes with the top three energy entropy ratios to form the frequency domain features, match the frequency domain sparsity of the electrical variables, and define as the frequency domain feature fusion vector; The calculation formula of the multi-scale time series feature F extracted by the multi-scale feature pyramid is: , wherein, is the transient feature vector, is the trend feature vector, is the frequency domain feature fusion vector; S4. Construct a deep gated adaptive network; Perform soft selection on the input features through the gated mechanism, retain key information by combining the memory vector, mix the ReLU and hyperbolic tangent activation functions in the residual block, adjust the nonlinear strength through learnable parameters, and perform channel-level scaling on the original input during the skip connection. Use a dynamic convolution kernel to process time series data, and its weight is jointly determined by the basic weight and the increment based on self-attention to achieve adaptive modeling of different time-span dependence relationships. The deep gated adaptive network includes the following modules: ① Gated feature selection module Perform a linear transformation on the input features through a learnable weight matrix and bias, generate a gated value between 0 and 1 through the Sigmoid function, multiply the feature vector element-wise by the gated value, and supplement the suppressed feature components with the memory vector; Feature gated vector The calculation formula is: , wherein, is the Sigmoid activation function; is the gated weight; F is the multi-scale time series feature extracted by the multi-scale feature pyramid; is the bias term; can control the retention and suppression of features, achieve soft selection, and adapt to the changes in feature importance under different working conditions; Gated output feature The calculation formula is: , wherein, is the Hadamard product; is a learnable memory vector to prevent the loss of weak features of urban pipe networks such as early fault signals; ② Hybrid activation residual block Apply ReLU and hyperbolic tangent activation to the gated output features simultaneously. The latter adjusts the output intensity through a Sigmoid gate. When performing residual connection, the original input is scaled by channels and added to the activated output. The scaling factor is a learnable parameter, expressed as: , wherein, is a hybrid activation function, which combines the ReLU activation function for processing linear features and the hyperbolic tangent function to capture non-linear oscillations, adapting to the complex non-linearity of electrical variables; z is the input of the hybrid activation function; is the ReLU activation function; is the hyperbolic tangent function; is the slope adjustment parameter of the activation function, which can be set to = 0.5; In the residual block, the input feature is the gated output feature. First, it is processed by a linear transformation and a hybrid activation function. At the same time, the original input is multiplied by a learnable scaling factor by channels, and then the two are added as the output, allowing the network to adaptively adjust the contribution intensity of the input feature and avoiding the vanishing gradient of the deep network; The output of the residual block is calculated by the formula: , wherein, is the weight matrix of the residual block; is the bias term of the residual block; represents a diagonal matrix; c is an adaptive scaling vector, representing the per-channel scaling factor of the input feature; ③ Dynamic convolutional time series processing module The convolutional kernel weight consists of a basic weight and a dynamic increment. The increment is adjusted by a self-attention mechanism. The attention weight calculates the mean of the autocorrelation matrix of all time-step feature vectors. During the convolution operation, the historical signals are weighted and summed according to the time span; The output of the time convolutional module is calculated by the formula: , wherein, is the time span of the convolutional kernel; is the dynamically adjusted convolutional kernel weight; is the value of the input signal at time step ; Dynamically adjusted convolutional kernel weights The calculation formula is as follows: , In the formula, is the base weight; is the dynamic weight increment; is the self-attention weight, which dynamically adjusts the convolutional kernel to enhance the modeling ability for long-range dependencies such as periodic loads and local mutations such as fault pulses; Self-attention weight The calculation formula is as follows: , In the formula, is the total length of the time series; is the feature vector of the input signal at time step t; is transpose of.

[0027] After cascading the output of the time convolution module and the output of the residual block, it is input into a preset Softmax classification function to output the predicted category; S5. Initialize the weights of the deep gated adaptive network; For the skewed characteristic of the electrical variable data distribution, determine the initialization range according to the characteristic distribution of the features, and statistically calculate the quartiles of each dimension of the multi-scale time series features extracted by the multi-scale feature pyramid. Use the difference between the first and third quartiles as a measure of the distribution dispersion, and generate a weight matrix with a uniform distribution accordingly, so that the initial weight range matches the actual data distribution and accelerates the model convergence; S501. Calculate the quartiles of each dimension of the multi-scale time series feature F extracted by the multi-scale feature pyramid is the first quartile, is the second quartile, is the third quartile; S502. Construct the initialization weight matrix The elements of satisfy a uniform distribution, and its range is determined by the data quantiles; The element in the i-th row and j-th column of the weight matrix , In the formula, represents the distribution symbol; represents a uniform distribution; is the lower bound scaling factor, which can be set to ; is the first quartile of the j-th dimension feature; is the third quartile of the j-th dimensional feature; b is the upper bound scaling factor, which can be set to b = 0.5; S6. Calculate the multi-objective joint total loss function of the depth gating adaptive network; The total loss function consists of a cross-entropy loss for focusing on correct classification, a focal loss for dynamically adjusting sample weights through predicted probabilities, and a maximum mean discrepancy loss for aligning the distributions of different classes in the feature space; the adjustment coefficient of the focal loss increases with the number of training epochs.

[0028] Total loss function The calculation formula is: , In the formula, is the weight of the cross-entropy loss; is the cross-entropy loss; is the focal loss, which makes the model focus on difficult examples by reducing the weights of easy-to-classify samples; is the weight coefficient of the maximum mean discrepancy loss, which is used to balance the relationship between distribution alignment and other loss terms, and can be set to = 0.2; is the maximum mean discrepancy loss, which is used to reduce the difference in feature distributions of different classes; The focal loss is used to enhance the model's attention to samples that are difficult to correctly classify. In specific implementation, first calculate the corresponding logarithmic loss value according to the predicted probability of each class by the model, and then weight it with a dynamic factor that decays with the prediction confidence. The more accurate the prediction of a sample, the smaller the weight given to its loss, while samples with inaccurate predictions obtain higher weights, thus prompting the model to focus on samples with blurred recognition boundaries or high noise during training; Focal loss The calculation formula is: , In the formula, c is a positive integer; is the total number of classes; is the predicted probability of the model for the c-th class; is the adjustment exponential factor; can dynamically control the model's attention to difficult-to-classify samples, and gradually increases with the number of training epochs. That is, in the initial stage of training, the model maintains a relatively balanced attention to all samples, which helps to learn the global structure. In the later stage of training, the model begins to focus on difficult examples and boundary samples, thereby enhancing the discriminative ability and robustness of the model; The calculation formula is: , In the formula, is the current training epoch; The maximum mean discrepancy loss is used to reduce the distribution difference between samples of the same class in the feature space while increasing the feature distance between different classes. In specific implementation, taking the binary classification task as an example, samples of different classes are input into the network and their deep feature representations are extracted. These features are mapped into a high-dimensional kernel space, and the feature means of each class in this space are calculated respectively. Finally, the distribution difference is measured by comparing the distance between the centers of these two classes. Maximum mean discrepancy loss The calculation formula is: , In the formula, is the norm in the reproducing kernel Hilbert space; is the number of samples of the first class; i is a positive integer; is the set of samples of the first class; is the kernel function mapping; represents the mapped feature of the i-th sample after passing through the hidden layer of the network; is the number of samples of the second class; j is a positive integer; is the set of samples of the second class; represents the mapped feature of the j-th sample after passing through the hidden layer of the network; In the multi-classification task, all classes can be combined pairwise, and the maximum mean discrepancy loss between each pair of classes is calculated. These loss values are weighted and summed as the final maximum mean discrepancy loss; S7. Iterative training and parameter update of the deep gated adaptive network to obtain the trained deep gated adaptive network; The gradient descent method is used for iterative training and parameter update, combined with the momentum mechanism to accelerate convergence and avoid local optima. At the same time, the early stopping mechanism is combined to prevent overfitting, ensuring that the model can efficiently learn the complex features of electrical variable data and maintain good generalization ability during training; S701. Define the parameter set of the model as , including the learnable parameters except the dynamically adjusted convolutional kernel weights ; S702. In each iteration process, by calculating the gradient of the total loss function with respect to the parameter , update the model parameters; The model parameters at the -th iteration are calculated as: , In the formula, represents the model parameters at the t-th iteration; is the learning rate, which can be set to = 0.001; is the total loss function for the t-th iteration For the parameter Gradient; For the Momentum term at the t-th iteration; S703. To accelerate the training process and avoid getting stuck in local optima, combined with the momentum mechanism, the momentum term provides inertia for parameter updates by accumulating historical gradient information, and the update method is expressed as: , In the formula, For the Momentum term at the t-th iteration; Is the momentum coefficient, used to control the decay rate of historical gradient information, and can be set to = 0.9; Is the learning rate; Is the total loss function for the t-th iteration For the parameter Gradient; S704. To prevent the model from overfitting, combined with the early stopping mechanism, during the training process, regularly calculate the loss value of the model on the validation set. If the decrease in the total loss function of the validation set is less than the preset threshold in 10 consecutive iterations, stop the training and retain the current model parameters as the final model; S8. Collect new urban pipe network operation status data, preprocess the new urban pipe network operation status data to obtain preprocessed new urban pipe network operation status data, ensure that the input features are consistent with the training set in terms of data distribution and expression ability, so as to ensure the reliability and consistency of the inference results; input the preprocessed new urban pipe network operation status data into the trained deep gated adaptive network, and the trained deep gated adaptive network analyzes the preprocessed new urban pipe network operation status data and then outputs the status label corresponding to the new urban pipe network operation status data to realize the identification of the urban pipe network operation status.

[0029] As Figure 4 shown, analyze the robustness of different methods in a noisy environment, and observe the change trend of the classification accuracy of each method by systematically increasing the noise level. Experimental data show that as the noise intensity increases, the performance of all methods will decline, but the decline of this technology is the gentlest, and it still maintains good recognition ability under strong noise conditions, indicating that the multi-level anti-interference design of the present invention plays a role, the front-end adaptive noise reduction processing effectively suppresses noise interference, the multi-scale feature extraction enhances the noise resistance of features, and the memory vector and dynamic weight mechanism in the deep network also improve the noise tolerance of the model.

Claims

1. A method for identifying the operating state of urban pipe networks, characterized in that, Including: S1. Collect multiple groups of urban pipe network operation status data and perform status annotation on each group of urban pipe network operation status data to obtain multiple groups of annotated urban pipe network operation status data. The multiple groups of annotated urban pipe network operation status data together constitute a training dataset; The urban pipe network operation status data includes the measured values of electrical variables at multiple key nodes in the urban pipe network; S3. Preprocess the multiple groups of annotated urban pipe network operation status data; S4. Extract features through a multi-scale feature pyramid; S5. Construct a deep gated adaptive network; Soft-select the input features through a gating mechanism, and retain key information in combination with the memory vector; Mix the ReLU and hyperbolic tangent activation functions in the residual block, adjust the non-linear intensity through learnable parameters, and perform channel-level scaling on the original input during skip connection; Adopt a dynamic convolution kernel to process time series data; S8. Initialize the weights of the deep gated adaptive network; S9. Calculate the multi-objective joint total loss function of the deep gated adaptive network; The total loss function consists of a cross-entropy loss for focusing on correct classification, a focal loss for dynamically adjusting the sample weights through the predicted probabilities, and a maximum mean discrepancy loss for aligning the distributions of different classes in the feature space; S11. Iteratively train and update the parameters of the deep gated adaptive network to obtain a trained deep gated adaptive network; S12. Collect new urban pipe network operation status data, preprocess the new urban pipe network operation status data to obtain preprocessed new urban pipe network operation status data, input the preprocessed new urban pipe network operation status data into the trained deep gated adaptive network, and the trained deep gated adaptive network analyzes the preprocessed new urban pipe network operation status data and outputs the status label corresponding to the new urban pipe network operation status data.

2. The urban pipe network operation state recognition method according to claim 1, characterized in that, The method for obtaining the measured values of electrical variables at multiple key nodes in the urban pipe network in S1 includes: S101. Deploy voltage-current composite sensors at key nodes; S102. The voltage-current composite sensor discretely samples the sensed analog electrical signal through an analog-to-digital conversion module to obtain the sampled electrical signal. The sampling frequency is dynamically adjusted according to the working conditions, including: S1021. Input the analog electrical signal; S1022. Condition the front-end signal to obtain the processed analog signal; S1023. Input the processed analog signal into the analog-to-digital conversion module, and the analog-to-digital conversion module internally converts the continuous analog waveform into 24-bit discrete digital values; S103. The sampled electrical signal is channel numbered according to its corresponding measurement node.

3. The method for identifying the operation state of an urban pipe network according to claim 1, characterized in that The status labels for status annotation in S1 include "fault" and "normal operation".

4. The urban pipe network operation state recognition method according to claim 1, characterized in that The preprocessing in S2 includes: S201. Normalize the multiple groups of annotated urban pipe network operation status data to obtain normalized annotated urban pipe network operation status data; S202. Denoise the normalized annotated urban pipe network operation status data.

5. A method for identifying the operating state of an urban pipe network according to claim 1, characterized in that, The feature extraction through a multi-scale feature pyramid in S3 includes: S301. Perform empirical mode decomposition on the denoised signal, extract the modal components reflecting the transient oscillation characteristics, calculate the instantaneous energy of each component through Hilbert transform, and extract multiple statistics to form a transient feature vector; S302. Use the dynamic time warping algorithm to align all sample signals with the reference mean sequence, extract multiple trend features from the optimal matching path, and perform wavelet packet decomposition; S303. Select the frequency band node with the highest energy entropy ratio to reconstruct the frequency domain signal, extract the frequency domain features, and splice the three types of features to form a multi-scale pyramid.

6. The urban pipe network operation state recognition method according to claim 5, wherein, The S301 performing empirical mode decomposition on the denoised signal, extracting the modal components reflecting the transient oscillation characteristics, calculating the instantaneous energy of each component through Hilbert transform, and extracting multiple statistics to form a transient feature vector includes: The measured value of the i-th channel of the electrical variable at time t after wavelet denoising processing is subjected to empirical mode decomposition to obtain modal components; S3012. Calculate the Hilbert marginal spectrum energy of each modal component; S3013. Construct a transient feature vector.

7. A method for identifying the operating state of an urban pipe network according to claim 5, characterized in that, The S303 selecting the frequency band node with the highest energy entropy ratio to reconstruct the frequency domain signal, extracting the frequency domain features, and splicing the three types of features to form a multi-scale pyramid to achieve the full-scale characterization from microscopic transient to macroscopic trend includes: S3031. Perform multi-layer wavelet packet decomposition on the signal, calculate the weighted sum of the proportion of the energy of each node in the total energy and its frequency band information entropy as the node weight coefficient, select the three nodes with the highest weights, reconstruct the frequency domain component with their basis functions and weight coefficients, extract the relevant features, and splice the transient, trend, and frequency domain features into a multi-scale feature pyramid; S3032. Select the nodes with the top three energy entropy ratios to form the frequency domain features and match the frequency domain sparsity of the electrical variables.

8. A method for identifying the operation state of an urban pipe network according to claim 1, characterized in that, The S5 initializing the weights of the deep gated adaptive network includes: S501. Calculate the quartiles of each dimension of the multi-scale time series features extracted by the multi-scale feature pyramid; S502. Construct an initial weight matrix whose elements satisfy a uniform distribution.

9. The method for identifying the operation state of an urban pipe network according to claim 1, wherein The total loss function described in S6 has the following calculation formula: ， Wherein, is the weight of the cross-entropy loss; is the cross-entropy loss; is the focal loss; is the weight coefficient of the maximum mean discrepancy loss; is the maximum mean discrepancy loss.

10. The method for identifying the operation state of an urban pipe network according to claim 1, characterized in that, The S7 iterative training and parameter update of the deep gated adaptive network includes: S701. Define the parameter set of the model as ; S702. In each iteration, by calculating the total loss function for the parameters gradient , update the model parameters; S703. Combine the momentum mechanism, and the momentum term provides inertia for parameter update by accumulating historical gradient information.

Images